betaFeedback

WorksPhysics and Astronomy

Precision-informed Fourier compression of refractive index tomograms with voxel-wise error control

Abstract

Significance: Holotomography (HT) generates large refractive index (RI) volumes and time series, creating a need for compact storage with controlled numerical fidelity. Aim: We sought to reduce storage by directly encoding reconstructed three-dimensional RI volumes in Fourier space with voxelwise validation against an assigned error budget. Approach: Quantized Fourier coefficients represent volumetric structure, while spatial residual quantization uses a 1% safety margin without storing separate corrections for individual voxels. We evaluated cellular and organoid data, concurrent processing, and extensions to dielectric-tensor and absorption tomography. Results: At a budget of 0.0001 RI units, complete packets achieved compression ratios of 7.89-9.50 relative to float32 arrays and were 28-30% smaller than BitRound-Pcodec packets satisfying the same error limit. Concurrent processing reduced encoding and decoding wall times by factors of 5.14 and 4.81 without changing packets. The fixed margin passed all primary-budget tests but failed at one smaller budget in both precision-sweep inputs. Conclusions: Precision-informed volumetric Fourier coding reduces storage of holotomograms with verified voxel-wise fidelity at the primary budget, at additional computational cost.

Statement of Discovery: In holotomography (HT), direct encoding of reconstructed three-dimensional refractive index (RI) volumes in Fourier space reduces storage with a residual safety margin and voxelwise error validation. The tested primary setting satisfies a 10−410^{-4} RI-unit tolerance without storing additional voxel replacement values. The framework also accommodates dielectric-tensor and absorption tomograms through quantity-specific error criteria.

Introduction

Holotomography (HT) reconstructs three-dimensional refractive index (RI) distributions for label-free analysis of cells and tissues[1, 2]. High-resolution RI imaging generates large datasets that must be stored and recovered for quantitative analysis. Label-free contrast enables repeated observation over extended periods, including large area tissue imaging, long-term organoid imaging, and four-dimensional measurements of intracellular dynamics[3, 4, 5, 6, 7, 8]. Storing a volume at each time point expands the data burden as imaging progresses from individual cells to organoids and tissues. Faster scanning and illumination multiplexing[9, 10] further increase the amount of RI data to store, transfer and recover.

RI compression must reduce stored bytes while limiting changes to the quantitative voxel values. Chunked formats such as OME-Zarr support scalable access to multidimensional bioimaging arrays[11, 12], while the numerical representation and codec determine each chunk's size and recovery cost. Lee et al. established HT compression baselines and showed that compression ratio and processing bandwidth must be considered alongside transfer bandwidth[13]. Bit Grooming reduces stored precision to improve lossless compressibility[14]; precision reduction followed by numerical compression, as in BitRound–Pcodec, provides a direct storage-saving pipeline[15, 16]. Digit Rounding similarly reduces mantissa precision before lossless coding but preserves significant decimal digits rather than a fixed absolute RI-error tolerance[17]. Encoding correlations across all three spatial dimensions offers a complementary way to reduce storage while controlling each voxel's numerical error.

A compression tolerance expressed in RI units limits the numerical change introduced into a reconstruction, separately from the accuracy of the original measurement. Metrological assessment of HT distinguishes reconstructed RI fidelity from spatial resolution and other imaging characteristics[18]. Where independently characterized, system precision can inform the allowed change, while compression error is evaluated relative to the supplied reconstruction. Measurement uncertainty has similarly motivated precision reduction in cryo-electron microscopy and tomography[19].

Scientific compressors already control numerical error through prediction, transforms and multilevel coding[20]. SZ combines prediction with error-controlled quantization, while zfp transforms floating-point arrays[21, 22, 23]; tensor decomposition[24] and MGARD[25] provide other multidimensional representations. SPERR combines wavelet coding with pointwise corrections[26], and the FFCz preprint considers joint spatial- and frequency-domain constraints[27]. Storage comparisons must count coefficients, residuals, voxel replacement data and metadata together at a verified voxelwise tolerance. Analysis of zfp further shows how errors at individual numerical stages propagate to its mode-specific bounds[28].

We directly encode reconstructed RI tomograms in three-dimensional Fourier space and validate decoded voxels against an absolute RI-error tolerance. Quantized Fourier coefficients represent volumetric structure, and spatial residuals correct local errors using a quantization interval of 1.98ε1.98\varepsilon without storing an additional list of voxel indices and replacement values for errors exceeding the tolerance. We evaluate complete-packet storage and computational cost against local BitRound–Pcodec and zfp implementations on public HeLa and organoid data under matched fidelity criteria. Experiments varying coefficient spacing examine how storage is divided between coefficients and residuals and the numerical limits of the fixed margin, while concurrent processing tests time-lapse throughput with unchanged independent packets. Extensions to incoherent dielectric tensor tomography (iDTT)[29] and quantitative absorption tomography (QAT) assess component-specific coding and error criteria[30].

Methods

Reconstructed Tomography Data and Sampling

The compression inputs were reconstructed RI tomograms. HT reconstruction rests on inverse-scattering theory and diffraction-based backpropagation[31, 32], using either white-light illumination with axially scanned image stacks[33, 34, 35] or coherent illumination with a scanned incidence angle[36, 37].

Public HeLa and organoid RI arrays were obtained from the BioImage Archive dataset S-BIAD1739 associated with the HT compression benchmark of Lee et al. 13,38 (Supplementary Methods S1). The saved HeLa and organoid arrays used as native references in this study were float64. We retained those arrays as the numerical references and used their float32 casts as working inputs. Each working subvolume contained 2,097,152 voxels, corresponding to 8 MiB. Supplementary Methods S10 specifies the reproduction materials and remaining validation limits.

For the time-lapse experiment, we extracted 16 consecutive organoid subvolumes at archive time indices 75–90, using the same spatial region at every time point. Each subvolume contained 32×256×25632 \times256 \times256 voxels. The spatial location was selected from the native morphology before timing and held fixed across all time points.

The extension experiments used one region of interest (ROI) from iDTT and one melanoma ROI from QAT[29, 30]. The iDTT and QAT data used in this study will be deposited in the BioImage Archive.

Compression Tolerance and Numerical Reference

Let n0(v)n_0(v) denote the native reconstructed RI at voxel vv, x(v)=f32(n0(v))x(v)=f_{32}(n_0(v)) the float32 working value, and n^(v)\hat{n}(v) the decoded RI. Here, f32f_{32} and f64f_{64} denote casting to float32 and float64, respectively. We evaluated the maximum absolute RI error against an assigned tolerance ε\varepsilon, accepting a setting only if

δnmax⁡=max⁡v∣n^(v)−n0(v)∣≤ε.\delta n_{\max}=\max_{v}\left|\hat{n}(v)-n_0(v)\right|\leq\varepsilon.

The primary ε=10−4\varepsilon=10^{-4} RI-unit setting was the compression tolerance adopted in this study, selected with reference to our experimental measurements using an HT-X1 system (Tomocube Inc., Republic of Korea). This tolerance is defined relative to the supplied reconstruction; it is not a claim of absolute RI accuracy or a universal instrument specification. Different physical quantities in the extension experiments used the separate criteria specified below. We chose a maximum-error criterion to limit every voxel, including rare local changes that an average error can conceal. Root-mean-square error (RMSE) is also reported as a complementary metric. An orthonormal transform preserves squared error in exact arithmetic, but small transform-domain RMSE alone does not ensure a voxelwise maximum-error bound.

Fourier Coefficient Representation and Lossless Coding

For a working volume xx containing NN voxels, the mean μ\mu and Fourier coefficients FF were computed in float64 as

μ=1N∑vx(v), F=Fr(f64(x)−μ),\mu=\frac{1}{N}\sum_{v}x(v),\ F=F_r\left(f_{64}(x)-\mu\right),

where FrF_r denotes an orthonormal three-dimensional real fast Fourier transform (FFT). The stored half-spectrum had shape (Nz,Ny,⌊Nx/2⌋+1)(N_z,N_y,\lfloor N_x/2\rfloor+1), where NzN_z, NyN_y and NxN_x are the voxel counts along the three array axes. At frequency-grid index kk, integer arrays qRq_{\mathrm{R}} and qIq_{\mathrm{I}} were obtained by independently quantizing the real and imaginary parts of FF with coefficient interval ΔF\Delta_F. The function rint rounds to the nearest integer with ties to even:

qR(k)=rint⁡(Re⁡F(k)ΔF),qI(k)=rint⁡(Im⁡F(k)ΔF).q_{\mathrm{R}}(k)=\operatorname{rint}\left(\frac{\operatorname{Re} F(k)}{\Delta_F}\right),\quad q_{\mathrm{I}}(k)=\operatorname{rint}\left(\frac{\operatorname{Im} F(k)}{\Delta_F}\right).

Quantized integer coefficients are converted to floating-point values for dequantization and inverse FFT computation. The inverse-transform prediction is computed in float64 and cast to float32 only after reconstruction and mean addition. The integer storage width is independent of the floating-point precision used for reconstruction. The primary HT setting was ΔF=5ε\Delta_F=5\varepsilon, an empirical rule selected during method development and fixed before the HeLa/organoid comparison. The factor 5 is an empirically chosen value held constant across the comparison, not a universal optimum or a chunk-size scaling law. Figure 4c,Figure 4d gives the tested size trade-off: the minima differ between HeLa and organoid. The orthonormal FFT convention avoids the linear-in-voxel-count amplitude factor of an unnormalized discrete Fourier transform, but packet efficiency can still depend on chunk shape and boundary structure.

Precision and storage, voxelwise error control, and packet composition

Figure 4. Precision, residual error control and packet composition. ε\varepsilon, assigned refractive index (RI)-error tolerance; ΔF\Delta_F, Fourier coefficient interval. (a) Complete-packet compression ratio at 14 logarithmically spaced tolerances per representative subvolume, from 5×10−35\times10^{-3} to 5×10−65\times10^{-6}, with ΔF=5ε\Delta_F=5\varepsilon and residual spacing 1.98ε1.98\varepsilon without voxel replacement. Red crosses identify measured settings exceeding the native-reference bound; they are not bound-compliant performance points. (b) Maximum native-reference RI error divided by ε=10−4\varepsilon=10^{-4} for Fourier prediction alone, conventional residual quantization at 2ε2\varepsilon and the adopted residual quantization at 1.98ε1.98\varepsilon. The two residual settings were tested independently. The dashed line marks the limit; the inset reproduces the near-boundary bar heights. Counts give violating HeLa/organoid voxels. (c,d) Coefficient, residual and header bytes for eight ΔF/ε\Delta_F/\varepsilon ratios at ε=10−4\varepsilon=10^{-4}; all 16 settings pass the fixed tolerance. Compression ratios use the same uncompressed float32 size as the reference.

Quantized coefficients were partitioned into four radial bands to form separate numerical coding streams. For per-axis FFT frequencies fzf_z, fyf_y, fxf_x in cycles per voxel, the normalized radius ρ(k)\rho(k) was

ρ(k)=2fz2+fy2+fx2.\rho(k)=2\sqrt{f_z^2+f_y^2+f_x^2}.

Boundaries at 0.25, 0.5 and 1.0 defined four bands, with equality assigned to the higher band. The metric uses per-axis Nyquist-normalized digital frequencies; it does not represent physical optical-transfer support. All half-spectrum positions were represented, and no Ewald-volume or resolution-based hard mask was applied. All four bands use the same coefficient interval ΔF\Delta_F.

Quantized real and imaginary arrays were cast to the smallest supported signed integer width, 16, 32 or 64 bits, covering their full ranges; all frequency bands used that same integer width. The integer storage width for the spatial residual was chosen separately. Each stream was losslessly encoded with Pcodec16 at level 6. The width is chosen over the full real or imaginary integer array and is used for all its frequency bands, rather than assigning 16, 32 and 64 bits to different radial intervals.

Spatial Residual with a Safety Margin and Decoding

The stored integers define reconstructed Fourier coefficients F~\widetilde{F} and a spatial prediction pp, where ii is the imaginary unit and Fr−1F_r^{-1} denotes the inverse real FFT:

F~(k)=ΔF[qR(k)+iqI(k)],p=f32(Fr−1(F~)+μ).\widetilde{F}(k)=\Delta_F\left[q_R(k)+i q_I(k)\right],\quad p=f32\left(F_r^{-1}(\widetilde{F})+\mu\right).

The encoder computed the integer spatial residual r(v)r(v) using a residual tolerance εr=0.99ε\varepsilon_r=0.99\varepsilon and quantization interval 2εr=1.98ε2\varepsilon_r=1.98\varepsilon. The decoded volume, denoted y0y_0 in the expression below, was formed directly from this residual and the prediction:

r(v)=rint⁡(f64(x(v))−f64(p(v))1.98ε),r(v)=\operatorname{rint}\left(\frac{f64(x(v))-f64(p(v))}{1.98\varepsilon}\right),
y0(v)=f32[f64(p(v))+1.98εf64(r(v))].y_0(v)=f32\left[f64(p(v))+1.98\varepsilon f64(r(v))\right].

In exact arithmetic, nearest-integer quantization at spacing 1.98ε1.98\varepsilon limits the residual error relative to the working input to 0.99ε0.99\varepsilon. The remaining 0.01ε0.01\varepsilon provides headroom for working-array and final-output casts; it is not a universal floating-point error bound. Float64 subtraction and residual addition reduce intermediate cancellation and rounding before the required float32 output cast; they do not add information to the input. The final decoded RI is the residual-corrected output y0y_0.

Every reported error was measured against the native reference using full-volume float64 differences. All primary HT settings passed at ε=10−4\varepsilon=10^{-4}. The precision sweep identified two failing settings at a smaller tolerance (Figure 4a); these are explicitly marked and excluded from bound-compliant performance claims. Supplementary Methods S1 and S3 describe the numerical limitations. Paired measurements of the residual-margin change are described in Supplementary Methods S7.

Each packet for a single field contains the mean, coefficient streams, residual and metadata needed for decoding without the original volume. An additional voxel replacement step (a guard), when used, stores voxel indices and source values to replace decoded values that fail the error check. Complete packet sizes include all headers, compressed arrays, any enclosing container, and replacement data required by a comparator or the joint tensor error limits. Packet framing is documented with the reproduction materials in Supplementary Methods S10.

Reference Codecs and Parameter Selection

BitRound–Pcodec and zfp were evaluated on the same float32 working inputs. BitRound used the Numcodecs mantissa-rounding implementation[15, 16] followed by Pcodec at level 6. zfp provides a compressed representation of multidimensional floating-point arrays[23, 39]; we used zfpy in fixed-accuracy mode with its header enabled and verified the final error against the native reference.

For each comparator, we used the tested setting that produced the smallest complete packet while keeping the RI error within the specified limit at every voxel. We reused these settings and the Fourier coefficient quantization rule in the residual-margin experiments without further tuning. The main figure therefore compares these fixed settings; other settings may yield smaller packets. Supplementary Methods S2 lists the tested parameter values and reports two additional comparisons: one that stores replacement values for voxels exceeding the error limit, and another that compares codecs at similar measured maximum errors. Supplementary Methods S3 reports the precision and coefficient–correction experiments. Supplementary Methods S11 reports packet sizes for FixedScaleOffset–Pcodec at levels 6 and 12; its encoding and decoding times were not measured.

Metrics, Verification and Figure Preparation

Compression ratio was Bref/BpacketB_{\mathrm{ref}} / B_{\mathrm{packet}}, where BrefB_{\mathrm{ref}} is the uncompressed reference size and BpacketB_{\mathrm{packet}} includes all headers and payloads; storage savings relative to a comparator were 1−Bproposed/Bcomparator1 - B_{\mathrm{proposed}} / B_{\mathrm{comparator}}, using the complete packet sizes BproposedB_{\mathrm{proposed}} and BcomparatorB_{\mathrm{comparator}}. Reference sizes used 4 bytes per scalar voxel, six unique float32 fields for iDTT and two float32 fields for QAT.

Throughput was reported in MiB/s. Independent decoding checks evaluated maximum absolute error, RMSE and the number of bound violations over full volumes. Native and decoded images shared spatial locations and display limits. Complete metric definitions, verification procedures and figure-selection rules are provided in Supplementary Methods S8.

Timing, Throughput and Memory Measurements

All codecs were timed on an Apple M4 Max host with 16 logical CPUs and 64 GiB of memory, using one FFT worker per job. Static HT and extension timings covered complete encoding or decoding, including required encoder-side reconstruction and any voxel replacement steps required by the comparator codecs or the tensor error limits, while excluding parameter screening, file I/O and separate quality validation. Time-lapse batch timing included warm local reads and buffered packet writes. Timings used medians from 5 static HT, 9 time-lapse or 21 extension repetitions at each setting. Extension peak resident set size (RSS) was measured separately in fresh processes. Software versions, warmups, timing order and memory protocols are provided in Supplementary Methods S4 and S6.

Bounded Concurrent Processing of Time-lapse Volumes

Each time point was encoded separately. Both Fourier and BitRound–Pcodec used a thread pool to process up to WW volumes simultaneously, and results were collected in input order. Figure 3 shows how performance changes with WW for each codec. The horizontal BitRound–Pcodec line shows its serial performance (W=1W = 1). Comparisons between the two codecs at the same WW, together with timing and memory measurement methods, are reported in Supplementary Methods S4.

Bounded concurrent processing of a time-lapse region of interest

Figure 3. Bounded concurrent processing of a time-lapse region of interest (ROI) at unchanged compression ratio. Fast Fourier transform (FFT); W, concurrent-volume limit. (a) Native sections of the same organoid ROI at t=75t = 75, 80, 85 and 90, corresponding to 0, 50, 100 and 150 min relative to t75t_{75}. Images use the same plane, crop and linear grayscale limits. Scale bars, 10 μm10\,\mu\mathrm{m}. Timing uses the complete 32×256×25632\times256\times256 subvolume at each of 16 distinct time points t=75t = 75–90; the schematic illustrates independent-volume scheduling. (b,c) Encoding and decoding wall times at every integer WW from 1 to 16, with one FFT worker per Fourier volume. Upper plots show total time for 16 volumes; lower plots show total/16, rather than task latency. Points are medians of nine repeats; error bars span minima to maxima. Fourier uses ΔF=5ε\Delta F = 5\varepsilon and residual spacing 1.98ε1.98\varepsilon without additional voxel replacements, with ε=10−4\varepsilon= 10^{-4}. Packets are identical across WW, compression is 8.59×8.59\times and every decoded voxel stays within the error tolerance relative to its native value. Dashed gray lines show BitRound–Pcodec. Solid gray horizontal lines show the local tolerance-matched BitRound–Pcodec W1 median held fixed across the axis, using the method family of [13].

Tensor and Absorption Tomography

For iDTT, the six unique components of the symmetric relative-permittivity tensor ϵ\epsilon were encoded independently, and the full tensor was restored by symmetric duplication. Its isotropic part mm and deviatoric part DD were defined by m=tr⁡(ϵ)3m = \frac{\operatorname{tr}(\epsilon)}{3}, D=ϵ−mID = \epsilon- mI, where tr⁡\operatorname{tr} is the trace and II is the 3×33 \times3 identity matrix. All codecs satisfied ∣δm(v)∣≤10−4\lvert\delta m(v) \rvert\le10^{-4} and ∥δD(v)∥F≤3×10−5\lVert\delta D(v) \rVert_F \le3 \times10^{-5} at every voxel, where δ\delta denotes the compression-induced change and ∥⋅∥F\lVert\cdot\rVert_F is the Frobenius norm. After decoding all components, the same final check was applied to every codec. At voxels that failed either tensor error check, all six values were replaced with their stored source values. For the Fourier method, this joint check followed residual reconstruction without separate replacements of individual component values. Compression ratios used the six unique float32 fields as the uncompressed reference. The m/Dm/D decomposition defines error metrics here; the packet still stores six unique tensor components, not seven m/Dm/D fields. In a separate m/Dm/D representation, trace⁡(D)=0\operatorname{trace}(D)=0 leaves only five independent deviatoric components, plus mm. Historical representation tests in Supplementary Methods S5 use different error limits applied separately to each tensor component.

For QAT, we encoded the co-registered RI nn and extinction coefficient κ\kappa volumes from the 520 nm melanoma reconstruction. The constant medium RI nmn_m was subtracted from nn before encoding and added back after decoding. The absolute error tolerances were εn=10−4\varepsilon_n=10^{-4} for nn and εκ=10−5\varepsilon_\kappa=10^{-5} for κ\kappa. RI contrast used Fourier coding and spatial residual quantization; κ\kappa used an FFT-free mean-plus-spatial-residual representation. Both proposed fields used spacing 1.98 times their respective tolerances without additional voxel replacement; the comparator codecs retained a separate replacement step for each field. Data definitions, fixed settings and validation are provided in Supplementary Methods S5. The physical definition is the imaginary part of the complex RI n~=n+iκ\tilde{n}=n+i\kappa, with absorption coefficient μa=2k0κ\mu_a=2k_0\kappa and k0=2π/λk_0=2\pi/\lambda. The supplied reconstruction stores this quantity as −Vimag-V_{\mathrm{imag}}.

Results

Precision-informed Fourier Quantization with Voxelwise Error Control

The algorithm compresses a three-dimensional RI volume by quantizing its Fourier coefficients, then corrects local errors using a spatial residual with a 1% safety margin (Figure 1). Fourier coefficients describe RI variations at different spatial frequencies. Quantization maps nearby coefficient values to shared integer levels and can map weak coefficients to zero, reducing the information needed for lossless coding. The main HT tolerance was ε=10−4\varepsilon=10^{-4} RI units, selected with reference to our experimental measurements using the HT-X1 system (see Methods); final compliance was checked against the native reconstruction.

Precision-informed Fourier compression with a residual safety margin

Figure 1. Precision-informed Fourier compression with a residual safety margin. Refractive index (RI); fast Fourier transform (FFT); ϵ\epsilon, assigned absolute RI-error tolerance. (a) The study-selected RI-error tolerance is ϵ\epsilon; the residual tolerance is ϵr=0.99ϵ\epsilon_{r} = 0.99\epsilon, giving spacing 2ϵr=1.98ϵ2\epsilon_{r} = 1.98\epsilon. (b) A native HeLa subvolume is cast to float32, mean-centered and transformed by a three-dimensional orthonormal real FFT. Real and imaginary coefficients are quantized separately at ΔF=5ϵ\Delta_{F} = 5\epsilon and losslessly coded. Spectra show the kz=0k_{z} = 0, kx≥0k_{x} \ge0 half-plane with shifted kyk_{y} and a common −80-80 to 00 dB amplitude reference. (c) Stored coefficients generate the encoder-side prediction; a spatial integer residual is computed from the float32 working input without storing additional voxel replacement values. (d) Coefficients, residual and metadata form a self-contained packet. The restored RI and signed error are shown with an unsmoothed measured error profile below. All 2,097,152 voxels satisfy ϵ=10−4\epsilon= 10^{-4} in this example. Paired RI images share display limits. Scale bar, 10 μ\mum. Smaller-tolerance limitations are shown in Figure 4.

Each complete scalar packet contains the quantized Fourier coefficients, spatial residual and metadata needed to decode the volume without the original tomogram. Real and imaginary

Fourier coefficients of the mean-centered float32 volume were quantized and losslessly encoded with Pcodec. The coefficients defined the Fourier prediction, and an integer residual with spacing 1.98ε1.98\varepsilon produced the decoded float32 values. No additional voxel indices or replacement values were stored.

Spatial residual quantization reduced the representative HeLa maximum RI error from 1.64×10−31.64 \times10^{-3} after Fourier quantization to 9.90×10−59.90 \times10^{-5}, or 0.99ε0.99\varepsilon, without additional voxel replacements (Figure 1). All 2,097,152 voxels satisfied ε=10−4\varepsilon= 10^{-4}, and the decoded root-mean-square RI error was 5.72×10−55.72 \times10^{-5}. An independent decoder verified the full three-dimensional subvolume; Supplementary Methods S9 explains how compliance was checked before rounding.

Smaller Packets for Cellular and Organoid Refractive Index Tomograms

Fourier coding reduced complete packet sizes in both public RI datasets while satisfying the 10−410^{-4} RI bound (Figure 2). The comparison used 16 subvolumes per dataset, each containing 32×256×25632 \times256 \times256 voxels, with compression ratios measured against the same float32 reference size.

Storage efficiency and single-volume processing cost with a residual safety margin

Figure 2. Storage efficiency and single-volume processing cost with a residual safety margin. (a,b) Native and Fourier-decoded refractive index (RI) sections of HeLa cells and organoids, with matched 64×6464 \times64-pixel details and identical display limits within each specimen pair. Scale bars, 10 μm in overviews and 2 μm in details. The organoid illustration is an additional cell-rich chunk excluded from the aggregate bars. (c–e) Complete-packet compression ratio, encoding throughput and decoding throughput for 16 benchmark subvolumes per dataset, each containing 32×256×25632 \times256 \times256 voxels. Ratios use a common float32 reference size. Fourier coding uses ΔF=5ε\Delta_F = 5\varepsilon and residual spacing 1.98ε1.98\varepsilon without additional voxel replacements, with ε=10−4\varepsilon= 10^{-4}.

Fourier compression achieved aggregate ratios of 7.89 for HeLa and 9.50 for organoid data, compared with 5.55 and 6.80 for BitRound–Pcodec[13, 16] and 3.90 and 4.29 for the zfp[23] configurations shown in Figure 2c. Fourier packets were consequently 29.6% and 28.4% smaller than BitRound–Pcodec packets, and 50.5% and 54.9% smaller than the displayed zfp packets. Native and decoded sections, shown with matched display limits and enlarged regions, retained the visible cellular structures (Figure 2a,Figure 2b).

Fourier packets remained smaller than BitRound–Pcodec packets at nearly the same realized maximum RI error. BitRound–Pcodec reached approximately 6.11×10−56.11 \times10^{-5} RI units, whereas the primary Fourier setting reached approximately 0.99×10−40.99 \times10^{-4}. Recomputing the residual-margin codec with its tolerance set to BitRound’s realized maximum error gave ratios of 6.67 and 8.08 for HeLa and organoid data, retaining 16.9% and 15.8% packet-size reductions with no violations. Fourier packets were also smaller in the additional comparator tests with voxel replacement; the settings and verification results are reported in Supplementary Methods S2.

Fourier encoding and decoding were slower than both comparator codecs for single volumes. Encoding throughput was 153–158 MiB/s, compared with 379–389 MiB/s for BitRound–Pcodec and 537–545 MiB/s for the displayed zfp configurations (Figure 2d). Decoding throughput was 337–341 MiB/s, compared with (2.13–2.47)×10³ MiB/s and 597–616 MiB/s, respectively (Figure 2e).

In the same comparison on identical inputs, replacing the earlier 2ε2\varepsilon residual followed by voxel replacement with the 1.98ε1.98\varepsilon residual reduced encoding time by 6.29% and 7.42% for HeLa and organoid data.

Concurrent Processing Reduces Time-lapse Wall Time without Changing Packets

Concurrent processing shortened the total encoding and decoding times for the 16-volume organoid RI series without changing the packets produced by serial processing (Figure 3). Consecutive subvolumes from the same spatial location at time indices 75–90 were encoded independently, with the concurrent-volume limit WW varied from 1 to 16. Fourier and BitRound–Pcodec retained compression ratios of 8.59 and 6.36, respectively, at every WW. Packet hashes were unchanged, and all Fourier-decoded voxels remained within the 10−410^{-4} RI-error tolerance.

Increasing WW from 1 to 16 reduced total Fourier encoding time from 0.886 to 0.172 s, a 5.14-fold speedup, and decoding time from 0.396 to 0.0823 s, a 4.81-fold speedup. The corresponding amortized encoding times were 55.4 and 10.8 ms per volume, and decoding times were 24.8 and 5.15 ms per volume (Figure 3b,Figure 3c). These values are batch wall times divided by 16. The lowest measured medians occurred at W=8W = 8 for encoding and W=9W = 9 for decoding; W=16W = 16 was not uniformly optimal.

Precision and Coefficient Spacing Determine Different Parts of the Storage Tradeoff

Reducing the target RI tolerance generally increased packet size in both representative subvolumes (Figure 4a). Across 14 logarithmically spaced tolerances from 5×10−35\times10^{-3} to 5×10−65\times10^{-6}, with ΔF=5ε\Delta F = 5\varepsilon, compression ratios fell from 843 to 3.87 for HeLa and from 847 to 4.22 for organoid data. Of the 28 settings, 26 satisfied their own native-reference tolerance. At ε=8.51×10−6\varepsilon= 8.51\times10^{-6}, both inputs failed: maximum error was 1.002ε1.002\varepsilon, with 607 and 592 violating voxels. Red crosses mark these noncompliant points. Passing again at 5×10−65 \times10^{-6} shows that the behavior cannot be summarized by a single lower tolerance threshold.

Figure 4b compares Fourier prediction alone with two separate residual-quantization settings.

Relative to conventional residual quantization at spacing 2ε2\varepsilon, the adopted spacing 1.98ε1.98\varepsilon reduced the maximum native-reference error to approximately 0.99003ε0.99003\varepsilon in both representative inputs, with zero violations and no additional voxel replacement data.

Coarser Fourier quantization did not continuously reduce total packet size because coefficient savings were offset by a larger spatial residual stream (Figure 4c,Figure 4d). At a fixed tolerance of 10−410^{-4}, increasing ΔF/ε\Delta_F / \varepsilon from 1 to 10 reduced HeLa coefficient storage from 1.02 to 0.282 MiB, while residual storage increased from 0.0329 to 0.708 MiB. The smallest tested complete HeLa packet occurred at ΔF/ε=5\Delta_F / \varepsilon= 5, giving a compression ratio of 8.26. The organoid minimum occurred at a ratio of 3, giving 9.77; the fixed ratio of 5 gave 9.69. All 16 fixed-tolerance settings passed.

Thus, total storage reflected a balance between coefficient and residual bytes, and the best tested spacing differed between inputs.

Extension to Tensor and Absorption Tomograms

The compression scheme also accommodated iDTT tensor components and paired QAT RI/absorption fields under error bounds tailored to their physical quantities[29, 30] (Figure 5). The tests used one symmetric dielectric-tensor ROI and one total variation-regularized melanoma ROI containing paired RI-contrast and extinction-coefficient fields. Each extension enforced error bounds on the physical quantities being stored.

Extension to tensor permittivity and quantitative absorption tomography: relative-permittivity slices, absorption and refractive-index slices, and compression, encoding and decoding throughput plots

Figure 5. Extension to tensor permittivity and quantitative absorption tomography (QAT). Incoherent dielectric tensor tomography (iDTT); refractive index (RI). (a) Original and reconstructed relative-permittivity component εxx\varepsilon_{xx} slices from iDTT. The colorbar reports relative permittivity, not RI. (b) Matched original and reconstructed extinction-coefficient κ\kappa and RI nn slices from the melanoma total variation reconstruction at 520 nm. RI is displayed as the supplied medium index, 1.336, plus Δn\Delta n. Scale bars, 5 μ\mum. (c–e) Compression ratio and encoding/decoding throughput for six unique float32 tensor fields and paired Δn/κ\Delta n/\kappa fields. Throughput derives from the median of 21 timed runs after three warmups; error bars span the transformed timing interquartile range. Reference sizes are 48 MiB for iDTT and 16 MiB for QAT. Proposed scalar residuals use spacing 1.98ε1.98\varepsilon without additional replacements of individual component values. The final step that replaces all six values at voxels failing either tensor error check is used for all iDTT methods to satisfy ∣δm∣≤10−4|\delta_m|\leq10^{-4} and ∥δD∥F≤3×10−5\lVert\delta\mathbf{D}\rVert_F\leq3\times10^{-5}. QAT tolerances are εn=10−4\varepsilon_n=10^{-4} and εκ=10−5\varepsilon_\kappa=10^{-5}; the comparators retain separate voxel replacement for each field. All evaluated volumes satisfy these error limits.

The tensor comparison used six unique components under common isotropic and deviatoric error bounds. These components reconstructed the full symmetric 3×33 \times3 tensor. Letting m=tr⁡(ϵ)/3m=\operatorname{tr}(\epsilon)/3 and D=ϵ−mID=\epsilon-mI, all codecs were required to satisfy voxelwise ∣δm∣≤10−4|\delta m|\le10^{-4} and ∥δD∥F≤3×10−5\lVert\delta D\rVert_F\le3\times10^{-5}, with the same final step that replaces all six values at voxels that fail either tensor error check. Compression ratios used the six float32 fields, totaling 48 MiB, so eliminating symmetric duplication did not itself inflate the reported gain.

For the tensor ROI, Fourier compression achieved a ratio of 7.89, compared with 4.17 for BitRound–Pcodec and 4.96 for zfp, reducing packet size by 47.1% and 37.1%, respectively (Figure 5c). The maximum error in mean relative permittivity and the maximum Frobenius norm of the deviatoric-tensor error were 1.48×10−51.48\times10^{-5} and 3.00×10−53.00\times10^{-5}. The maximum principal-RI error was 9.66×10−69.66\times10^{-6}, below 10−410^{-4}, and the maximum error among pairwise principal-index differences was 1.31×10−51.31\times10^{-5}.

Fourier tensor compression used more processing time and peak memory than BitRound–Pcodec and zfp. Encoding and decoding took 335 and 140 ms, compared with 157/23.3 ms for BitRound–Pcodec and 194/82.4 ms for zfp. Peak process RSS was 298/296 MiB for Fourier encoding/decoding, versus 273/224 MiB and 230/220 MiB for the comparators. All methods used the same final step that replaces all six tensor values at voxels that fail either error check.

QAT compression used separate native-reference error tolerances for the paired RI-contrast Δn\Delta n and extinction-coefficient κ\kappa fields. We used the total variation-regularized green channel data at 520 nm, with Δn=n−nmed\Delta n=n-n_{\mathrm{med}}, nmed=1.336n_{\mathrm{med}}=1.336, and κ=−Vimag\kappa=-V_{\mathrm{imag}}. The two co-registered fields were each 32×256×25632\times256\times256 voxels, giving a combined float32 size of 16 MiB. Separate native-reference tolerances of 10−410^{-4} for Δn\Delta n and 10−510^{-5} for κ\kappa were checked for all three codecs. Proposed fields used 1.98ϵ1.98\epsilon residuals without additional voxel replacement; both comparator codecs stored source values to replace decoded voxels that failed the corresponding field’s error check.

The QAT κ\kappa path used a constant float32 prediction formed from its float64 mean, followed by spatial residual quantization at 1.98×10−51.98\times10^{-5} and Pcodec coding. This avoids an FFT for κ\kappa while retaining its spatial variation in the residual. This choice followed tests in which all quantized κ\kappa Fourier coefficients were zero. The selected representation and coefficient spacing were held fixed for the residual-margin evaluation (Supplementary Methods S5).

The combined Fourier and mean-plus-residual QAT representation produced smaller packets than both comparators, with greater runtime and peak memory use in the tested melanoma ROI. The combined compression ratio was 12.6, compared with 4.81 for BitRound–Pcodec and 10.5 for zfp, corresponding to 61.7% and 16.4% smaller packets (Figure 5c). Maximum errors were 9.90×10−59.90 \times10^{-5} for Δn\Delta n and 9.90×10−69.90 \times10^{-6} for κ\kappa, with no violations. Encoding/decoding took 65.5/27.5 ms, compared with 50.0/8.05 ms for BitRound–Pcodec and 41.0/15.4 ms for zfp (Figure 5d,Figure 5e). Peak RSS was 176/158 MiB, compared with 169/93.9 MiB and 125/88.6 MiB. The QAT coefficient spacing and the representation used for each field were selected on this same ROI and kept unchanged for the margin experiment. Supplementary Methods S5 reports the separate RI-contrast and extinction-coefficient packet sizes and the common container overhead.

Discussion

Direct compression of the reconstructed three-dimensional RI volume reduced storage with every tested primary-setting voxel within its RI-error ceiling. The method combines a Fourier representation of the RI volume with spatial residual correction and native-reference verification. At 10−410^{-4} RI units, HeLa and organoid packets were approximately 28–30% smaller than BitRound–Pcodec packets; tightening the Fourier tolerance to BitRound's realized error retained a 15.8–16.9% reduction (Figure 2). The latter comparison shows that the observed storage gain persists when the methods use nearly the same realized maximum error. This gain comes with additional single-volume computation, making storage capacity and transfer bandwidth important factors in codec selection[13].

Hologram codecs store holograms or complex wavefields for reconstruction, whereas the present method directly stores the reconstructed three-dimensional RI field. These hologram-domain codecs[40] use approaches such as Fourier-coefficient reduction[41], nonuniform companding[42], bandwidth-aware image-plane coding[43], assessment through reconstructed RI volumes[44], and adaptive short-time Fourier quantization[45]. Here, the error criterion applies directly to the decoded RI voxels. Single-cell RI tomograms have also been encoded using threedimensional Zernike descriptors[46]. Related precedents include three-dimensional Fourier quantization for volume compression[47] and magnetic resonance imaging compression that limits signal-to-noise ratio loss by relating error variance to thermal noise[48]. These precedents motivate direct volumetric coding with an RI-unit fidelity constraint for quantitative HT storage.

A reconstructed RI volume can be described by Fourier coefficients that specify the amplitude and phase of RI variations at different spatial scales. The numerical precision retained in a stored RI volume can be reduced subject to a specified reconstruction-error tolerance. Quantizing the Fourier coefficients reduces this numerical precision by mapping nearby coefficient values to shared integer levels, which can be encoded efficiently. Here, our experimental measurements using the HT-X1 system informed the selected tolerance; the compression experiments assess numerical fidelity relative to the saved reconstruction. Since coefficient errors combine during the inverse transform, spatial residuals use a tighter tolerance and final native-reference checks establish compliance for each evaluated setting. The method thus stores the reconstructed three-dimensional RI field with fewer bits while keeping the compression error within the study-selected tolerance. In HT, accessible spatial bandwidth also determines direction-dependent optical resolution[49]. The present codec retains all half-spectrum positions rather than applying a hard optical-bandwidth mask.

Most of the storage reduction achieved came from quantization at the selected RI-error tolerance, followed by lossless compression. Direct spatial quantization already provided substantial compression under the same error limit. Fourier-domain quantization provided a further reduction in stored size of 9.3% for HeLa and 14.2% for organoid volumes under matched lossless compression settings (7.8% and 12.3% relative to Pcodec level 12; Supplementary Methods S11), supporting its use in the present codec. In these same inputs, 55.7% of HeLa and 70.5% of organoid quantized Fourier scalar coefficients were zero, while 89.2% and 92.7% of residual integers were −1, 0 or +1. Grouping the same coefficient integers by frequency band also reduced their encoded size (Supplementary Methods S12). These measurements support the compact coefficient-plus-residual representation for the tested inputs.

The resulting advantage reflects the combined effects of the data structure, quantization, and encoding strategy, and its magnitude may vary across datasets.

Compressing the reconstructed volume lets us control the error in the RI values used for analysis. The three-dimensional Fourier coefficients describe spatial variation in the quantitative RI field, and a full-volume native-reference check determines whether a selected setting meets the tolerance. Memmolo et al. encoded single-cell RI tomograms using three-dimensional Zernike descriptors[46]. General volume compression has previously used three-dimensional Fourier quantization[47], magnetic resonance imaging compression has linked error variance to thermal noise and image signal-to-noise ratio[48], and SPERR and the FFCz preprint provide transform-correction precedents[26, 27]. Our contribution is the integration and evaluation of volumetric Fourier coding with a study-selected RI tolerance and native-reference voxelwise validation in HT, including explicit tests of the margin’s numerical limits. This differs from compressing two-dimensional holographic wavefields and assessing their effects after RI reconstruction[44]. A bound on the decoded volume controls the stored quantitative output directly, without requiring hologram decoding and tomographic reconstruction to recover that output.

Compression efficiency depends on how much data are stored in the Fourier coefficients and how much are needed for spatial correction. Coarser quantization reduces the information stored in the Fourier coefficients but leaves larger deviations for the spatial residual to correct. Finer quantization preserves a more accurate Fourier prediction at the cost of additional coefficient storage. The minima observed in Figure 4c,Figure 4d identify an intermediate spacing at which these competing contributions yield the smallest complete packet. Their different locations show that the best tested balance depends on the input. Effective compression therefore requires choosing Fourier precision to minimize coefficient and correction storage together; the adopted residual margin controls quantization error and native-reference validation identifies compliant settings.

A residual tolerance of 0.99ε0.99\varepsilon eliminated the extra pass that identified voxels exceeding the single-field error limit and stored their indices and replacement values at the primary HT setting, reducing encoding time in the paired comparison. The 1% margin is an empirical numerical allowance. At ε=8.51×10−6\varepsilon= 8.51 \times10^{-6}, input and output casts together exhausted the margin and produced small violations in both precision-sweep inputs (Figure 4a). Deployment at another tolerance, numerical library or architecture therefore requires a check of every decoded voxel against the original reconstruction or a separately validated rounding strategy. The final iDTT joint physical correction remains necessary for the joint isotropic and deviatoric tensor error limits.

Bounded concurrent processing shortened the time needed to encode or decode independent volumes without changing the compressed data. Independent chunk processing is consistent with chunked storage workflows; the present timing experiment does not establish end-to-end OME-Zarr integration. Across the 16-volume organoid series, W1→W16W1 \to W16 reduced encoding and decoding wall times by factors of 5.14 and 4.81, while packet bytes, the compression ratio of 8.59 and decoded fidelity were unchanged (Figure 3). Limiting active volumes permits processing without holding the complete acquisition in memory. The lowest medians occurred at W8 and W9 for encoding and decoding, respectively. These results describe the tested host and finite batch; detailed scheduling behavior and matched-concurrency comparisons are given in Supplementary Methods S4.

The coding framework accommodated scalar RI volumes, iDTT and QAT, with error criteria tailored to the physical quantities being stored. For iDTT[29], separate tolerances constrained errors in mean permittivity and in the traceless part of the tensor that describes dielectric anisotropy. These trace- and Frobenius-based criteria are invariant under orthogonal changes of basis, although component-wise coding can yield basis-dependent packet sizes. Using the six independent tensor components as the uncompressed reference excluded gains from symmetric duplication. For QAT[30], the framework accommodated co-registered RI contrast and extinction coefficient κ\kappa under separate component-specific error tolerances. The RI contrast retained a volumetric Fourier representation, whereas all quantized Fourier coefficients of the extinction coefficient vanished at the selected spacing. The selected mean-plus-residual representation avoids this redundant transform. The spatial residual retained the extinction coefficient’s variation around its mean, so a constant prediction was sufficient for this field at the selected tolerance. Under the matched physical error criteria used in each experiment, the proposed method produced smaller complete packets than BitRound–Pcodec and zfp for both extensions, including a 16.4% reduction relative to zfp for the tested QAT ROI, at the cost of greater runtime and peak memory use (Figure 5). The κ\kappa result supports a mean-plus-residual representation at the selected tolerance; it does not demonstrate that Fourier coding benefits the absorption field.

Validation should extend beyond voxelwise reconstruction error to the preservation of quantitative analysis results. The bound limits numerical changes relative to the saved reconstruction and does not establish physical RI accuracy. Accuracy in derived quantities is a separate objective, addressed explicitly in multilevel scientific-data reduction[53]. Dry-mass metrology[54] and quantitative organoid segmentation[55] offer task-level tests of compressed RI data. Broader validation should include calibrated precision measurements and tests of integrated RI, segmentation and birefringence. Principal-axis accuracy also requires separate assessment near tensor eigenvalue degeneracy. Artificial intelligence-based segmentation, cell-type classification and virtual staining may tolerate compression errors larger than the 10−410^{-4} RI-unit tolerance used here. Characterizing the precision required by each task could enable more efficient compression under validated task-specific fidelity criteria. Such validation should encompass the complete analysis pipeline, because dry-mass precision depends on optical-path measurements, background estimation and cell segmentation[56].

Downstream validation should compare RI-derived lipid concentration, dry mass and threshold-based lipid-droplet segmentation between native and compressed volumes[57]. Lymphocyte-subtype classification provides a further endpoint based on morphological and biochemical features extracted from RI tomograms58. For virtual staining, validation should assess both inferred compartment masks and the dry-mass estimates obtained by applying those masks to quantitative phase data59.

Further work could test discrete cosine and wavelet transforms, assess the contribution of the Fourier transform and four-band partition, and explore frequency-dependent error allocation using calibrated noise estimates and voxelwise error validation. Future work may also test whether the observed compression gains translate to large-scale imaging workflows that include storage, transfer and analysis. Runtime and peak RSS were measured on one host, and whether decoding on different processor architectures meets the same error limits remains untested. The HTPack implementation stores a checksum of the decoded array and rejects a nonmatching reconstruction. A checksum mismatch identifies a difference from the certified decoded array; it does not by itself establish that the RI-error tolerance has been exceeded. Longer streaming benchmarks should include reads from uncached storage, data transfer, checks that decoding on different platforms meets the error limits and concurrency selection based on available CPU and memory resources. Integration with chunked bioimaging formats would allow evaluation of the complete storage-to-analysis workflow1.

Conclusion

Precision-informed three-dimensional Fourier coding reduced complete-packet storage for the tested RI tomograms under a verified 10−410^{-4} voxelwise error bound. Packets were 28–30% smaller than BitRound–Pcodec packets, with a 15.8–16.9% reduction retained after matching its realized maximum error. The fixed margin failed at one smaller tolerance in both precision-sweep inputs, so its validity requires numerical checks. Additional computation accompanied the primary-setting gain; concurrent processing improved batch throughput while preserving identical packets. Initial iDTT and QAT experiments extended the framework to component-specific error criteria.

Disclosures

Y.K.P. has financial interests in Tomocube, a company that commercializes HT instruments.

Juyeon Park receives financial support from Tomocube.

Code and Data Availability

The public HeLa and organoid RI datasets used for the static and time-lapse HT comparisons are available from the BioImage Archive under accession S-BIAD1739 (https://doi.org/10.6019/S-BIAD1739). The archive members, subvolume selection rules and ROIs used in this study are specified in the Methods and Supplementary Methods. The iDTT and QAT reconstruction data used for the extension experiments will be deposited in the BioImage Archive; accession numbers will be added when available. The research implementation and HTPack command-line interface will be released on GitHub.

Acknowledgments

This work was supported by the National Research Foundation of Korea (NRF), funded by the Ministry of Science and ICT (MSIT), Republic of Korea (Grant Nos. RS-2024-00442348 and RS-2022-NR068141); the Korea Institute for Advancement of Technology (KIAT) through the International Cooperative R&D Program (Grant No. P0028463); the Bio & Medical Technology Development Program of the NRF, funded by MSIT (Grant No. RS-2024-00440577); and the Commercialization Promotion Agency for R&D Outcomes (COMPA), funded by MSIT (Grant No. 2710110765). YongKeun Park also acknowledges support from the Merck fellowship.

We thank Hervé Hugonnet for his insightful comments and constructive feedback. OpenAI Codex was used to assist with language editing and manuscript formatting.

Author Contributions

YongKeun Park: conceptualization, methodology, supervision, funding acquisition, and writing—original draft. Dohyeon Lee: data curation, software, formal analysis, validation, visualization, and writing—review and editing. Juheon Lee: methodology, resources, validation, and writing—review and editing. Sehyeon Lee: resources, validation, and writing—review and editing. Juyeon Park: project administration, supervision, and writing—review and editing.

Author Biographies

Dohyeon Lee conducts research in biomedical optics at KAIST. His interests include holotomographic data processing, scientific data compression, and computational methods for managing large three-dimensional quantitative imaging datasets.

Juheon Lee conducts research in biomedical optics at KAIST. His research focuses on dielectric tensor tomography, polarization-resolved imaging, and quantitative three-dimensional characterization of optical anisotropy.

Sehyeon Lee conducts research in biomedical optics at KAIST. His interests include computational microscopy, quantitative absorption tomography, and reconstruction of three-dimensional optical properties from intensity measurements.

Juyeon Park conducts research in the Biomedical Optics Laboratory at KAIST. Her research focuses on developing label-free three-dimensional imaging and deep-learning methods for quantitative analysis of biological tissues and organoids. In particular, she investigates virtual staining of holotomographic images, with the goal of understanding cellular organization and dynamics without fluorescent labeling.

YongKeun Park is an endowed chair professor in the Department of Physics at KAIST and director of the KAIST Virtual 3D Biology Center. His research spans biomedical optics, wave physics, and holotomography, with a particular focus on developing label-free 3D imaging and quantitative analysis technologies for biological cells, tissues, and advanced 3D culture models. He is a fellow of Optica and SPIE.

Reference numbers correspond to the bibliography in the main article.

Data Selection and Error Reference

We used full-resolution HeLa and organoid refractive index (RI) arrays. Each dataset contributed 16 randomly selected subvolumes of 32×256×25632 \times256 \times256 voxels; incomplete boundary chunks were excluded. The input lists and source hashes identify the exact data used.

The saved HeLa and organoid float64 arrays were the error references, and their float32 casts were the working inputs. Residual spacing 1.98 ε1.98\,\varepsilon limits exact-arithmetic quantization error to 0.99 ε0.99\,\varepsilon. Input and output rounding also contribute to the final error, so we checked every decoded voxel against the native reference. Proposed single-field packets stored no extra voxel replacement values. The float32 working arrays and the 4-byte-per-voxel compression-ratio denominator are distinct from this native reference. We selected the primary holotomography (HT) tolerance of ε=10−4\varepsilon= 10^{-4} RI units with reference to our experimental measurements using an HT-X1 system (Tomocube Inc., Republic of Korea). The tests reported here establish compression fidelity relative to the saved reconstructions; they do not establish the absolute accuracy or measurement precision of those reconstructions.

Comparator Settings and Additional Comparisons

For BitRound, we tested 11-18 retained mantissa bits. For zfp, we tested tolerances from 2×10−52 \times10^{-5} to 1.6×10−31.6 \times10^{-3}. We selected the smallest total packet size for which every voxel met the 10−410^{-4} RI-error limit. The selected settings were 13 bits for BitRound and tolerance 4×10−44 \times10^{-4} for zfp; neither needed voxel replacement. These settings and the Fourier rule ΔF=5ε\Delta_F = 5\varepsilon were then kept fixed.

We also tested coarser BitRound and zfp settings with voxel replacement. Their packet sizes included both voxel indices and source values. At 12 retained bits, BitRound ratios were 4.12 and 4.56 for HeLa and organoid. At zfp tolerance 1.6×10−31.6 \times10^{-3}, ratios were 4.77 and 5.34. All met the native-reference error limit, but their packets were larger than the selected BitRound packets.

To compare methods at nearly the same measured maximum error, we set the Fourier tolerance to the selected BitRound maximum for each dataset, approximately 6.11×10−56.11 \times10^{-5}. Fourier ratios were 6.67 and 8.08, giving 16.9% and 15.8% smaller packets for HeLa and organoid. All 32 packets passed independent decoding. Parameter searches were excluded from timing; exact grids and selected values are retained in the benchmark records.

Precision and Fourier-spacing Experiments

Figure 4 used one fixed HeLa and one fixed organoid subvolume, selected by their native RI interquartile range. We tested 14 logarithmically spaced error tolerances from 5×10−35 \times10^{-3} to 5×10−65 \times10^{-6} with ΔF=5ε\Delta_F = 5\varepsilon. At approximately ε=8.51×10−6\varepsilon= 8.51 \times10^{-6}, both inputs failed: maximum errors were 1.002ε1.002\varepsilon, with 607 and 592 violating voxels. The other 26 settings passed, including the smaller endpoint.

Thus the tests do not define a single lower tolerance threshold below which all inputs fail. At the failing HeLa voxel, input casting, residual quantization and final output casting contributed approximately −5.95×10−8-5.95 \times10^{-8}, −8.412×10−6-8.412 \times10^{-6} and −5.14×10−8-5.14 \times10^{-8} RI units, respectively. Their sum exceeded the assigned tolerance because the combined cast errors exhausted the 8.51×10−88.51 \times10^{-8} safety margin. Alignment of the quantization grid with float32 values explains why a smaller tested tolerance could pass.

At ε=10−4\varepsilon= 10^{-4}, we tested coefficient-spacing ratios ΔF/ε\Delta_F / \varepsilon of 1, 2, 3, 4, 5, 6, 8 and 10. All 16 packets passed. Figure 4c,d separates coefficient, residual and header storage; the minima are the smallest packets among these tested settings. Figure 4b compares the Fourier prediction with two separately tested residual spacings, 2ε2\varepsilon and 1.98ε1.98\varepsilon.

The fixed rule ΔF=5ε\Delta_F = 5\varepsilon was chosen before the benchmark. All half-spectrum positions were retained. The four radial bands used the same coefficient spacing and integer width; separating bands changed only how values were grouped for lossless compression. Coefficients were stored in flattened C-order within each band.

Time-lapse Processing

Figure 3 used 16 consecutive organoid subvolumes from the same spatial region at archive time indices 75-90. Each volume was encoded independently. The scheduler kept at most WW jobs submitted but not yet consumed, collected results in input order and admitted the next job after consuming a result. The full series did not need to be held in memory.

Each codec, stage and concurrency setting was measured nine times after one warmup. Codec and concurrency order were reversed on alternating repetitions. Encoding included local input reads, float32 casting, compression and buffered packet writes. Decoding included packet reads and reconstruction. Downloading, decoded-array file writes and separate quality checks were excluded. Scheduler overhead was included.

Figure 3 shows batch wall time and that time divided by 16. The latter is amortized time per volume, not individual job latency. The fixed BitRound line is the local W=1W=1 median. BitRound remained faster at matched WW; at W=16W=16, its amortized encoding and decoding times were 4.52 and 0.796 ms per volume. Packet hashes and decoded arrays were unchanged across concurrency settings.

Tensor and Absorption Data

Dielectric-tensor Compression

The incoherent dielectric tensor tomography (iDTT) input was a real, symmetric relative-permittivity tensor[29]. We encoded six unique components, each containing 32×256×25632 \times256 \times256 float32 voxels, and restored the remaining components by symmetric duplication. Sampling was 0.0994 micrometers laterally and 0.4 micrometers axially. Compression ratios used the six unique fields, totaling 48 MiB. The input was the e array in 2_DT.mat from the 20260123/ps1um acquisition, with source-array crop [9:41, 222:478, 222:478, :, :]. Figure 5a displays the relative-permittivity component, whereas Figure 5b displays RI and extinction coefficient.

For tensor ϵ\epsilon, we defined the isotropic part mm and deviatoric part DD as

m=tr⁡(ϵ)3,D=ϵ−mI.m=\frac{\operatorname{tr}(\epsilon)}{3},\quad D=\epsilon-mI.

All codecs had to satisfy ∣δm(v)∣≤10−4|\delta m(v)|\leq10^{-4} and ∥δD(v)∥F≤3×10−5\lVert\delta D(v)\rVert_F\leq3\times10^{-5} at every voxel. With decoded-minus-native component errors eabe_{ab}, we evaluated

∥δD∥F2=∑a∈{x,y,z}(eaa−δm)2+2(exy2+exz2+eyz2).\lVert\delta D\rVert_F^2=\sum_{a\in\{x,y,z\}}(e_{aa}-\delta m)^2+2(e_{xy}^2+e_{xz}^2+e_{yz}^2).

After decoding, all codecs used the same joint check: if either tensor error reached or exceeded its limit within a 10−1210^{-12} margin, the packet supplied all six source values at that voxel. This affected 791 voxels for the proposed method. Fourier components used the 1.98 ϵ1.98\,\epsilon residual without separate component-wise replacement.

Diagonal and off-diagonal settings were selected before adopting the residual margin and then kept fixed. Selected Fourier scalar tolerances were 1.5×10−51.5\times10^{-5} and 10−510^{-5}; BitRound retained 16 and 4 bits; zfp tolerances were 4.8×10−44.8\times10^{-4} and 1.2×10−41.2\times10^{-4}. After decoding, we checked symmetry, positive eigenvalues, principal indices and pairwise index differences. These validation calculations were excluded from codec timing.

Earlier Tensor-representation Comparison

An earlier test used the previous codec with voxel replacement and a componentwise error limit of 10−410^{-4}. On the same region of interest (ROI), isotropic/deviatoric coding produced 12.3% smaller packets than six-component coding, but encoding and decoding were 2.4% and 4.8% slower and peak resident set size (RSS) increased by approximately 48 MiB. These results used different error criteria from Figure 5.

Refractive Index and Extinction-coefficient Compression

The quantitative absorption tomography (QAT) input was the co-registered melanoma RI and extinction-coefficient reconstruction at 520 nm, with total variation regularization and no aberration correction[30]. The two fields each contained 32×256×25632\times256\times256 float32 voxels, totaling 16 MiB, with isotropic sampling of 0.135 micrometers. The supplied medium RI was nmed=1.336n_{\mathrm{med}}=1.336. We subtracted it before encoding RI and added it back after decoding; the final RI error was checked against the native RI.

The selected error tolerances were εn=10−4\varepsilon_n = 10^{-4} and εκ=10−5\varepsilon_\kappa= 10^{-5}. BitRound retained 6 and 3 bits for RI contrast and extinction coefficient; zfp tolerances were 1.6×10−31.6 \times10^{-3} and 1.6×10−41.6 \times10^{-4}. Both comparators used voxel replacement where needed. Fourier settings were selected on the same ROI before adopting the residual margin and were not searched again.

The selected RI-contrast coefficient spacing was approximately 4.22×10−44.22 \times10^{-4}. For κ\kappa, all quantized Fourier coefficients were zero. We therefore omitted its fast Fourier transform (FFT) and used the native mean, calculated in float64 and cast to float32, as a constant prediction. Pcodec compressed the residual quantized at 1.98×10−51.98 \times10^{-5}, without extra voxel replacement.

The proposed RI-contrast packet occupied 0.862 MiB and the extinction-coefficient packet 0.412 MiB. Each includes its own field header and compressed streams; the common container added 661 bytes. The complete packet was 1.27 MiB, giving the combined ratio of 12.6 relative to the 16 MiB input. For BitRound–Pcodec, the RI-contrast and extinction-coefficient packets occupied 2.12 and 1.21 MiB; the corresponding zfp values were 1.2 and 0.322 MiB. The combined proposed result therefore includes Fourier coding of RI contrast and FFT-free mean-plus-residual coding of the extinction coefficient. For this ROI, the proposed extinction-coefficient packet was larger than the zfp packet; the smaller combined packet resulted from the reduction in RI-contrast storage.

The supplied reconstruction defines κ=−Vimag\kappa= -V_{\mathrm{imag}}, the negative imaginary reconstruction field. Figure 5b uses the same native and decoded arrays as the performance measurements, at the fixed plane z=10z = 10. Each pair shares linear grayscale limits. Negative reconstructed κ\kappa values were retained and are not interpreted as negative physical absorption. Exact source fields, crops and settings are recorded with the input manifests.

Timing and Memory Measurements

Experiments ran on an Apple M4 Max with 16 logical CPUs and 64 GiB of memory. Each Fourier job used one FFT worker. Tensor and QAT measurements used one outer job and one thread for OpenMP and BLAS. Software versions are recorded with the benchmarks.

Static HT measurements used one warmup and five encode/decode repetitions per input and setting. Tensor and QAT measurements used three warmups and 21 repetitions. Timings included complete codec operations: prediction, residual coding, headers and any required comparator or joint tensor replacement. Inputs were already in memory; file I/O, parameter searches and separate quality checks were excluded. Time-lapse timing included the local reads and packet handling described in S4.

Tensor and QAT peak RSS was the maximum of three fresh-process measurements, including inputs, outputs and runtime memory. Time-lapse RSS was sampled in the benchmark process. These two memory protocols were not treated as equivalent.

Effect of the Residual Safety Margin

We compared the previous residual spacing 2ε2\varepsilon followed by voxel replacement with the adopted spacing 1.98ε1.98\varepsilon without replacement. Both used identical inputs, Fourier coefficients and comparator settings.

At ε=10−4\varepsilon= 10^{-4}, the new margin reduced encoding time by 6.29% for HeLa and 7.42% for organoid, with zero violations. Packet size increased by 0.124% and decreased by 0.430%, respectively, because the new spacing changed residual compressibility. Simply removing replacement while retaining spacing 2ε2\varepsilon left 586 HeLa violations, with maximum error 1.001×10−41.001 \times10^{-4}.

Metrics and Independent Verification

Compression ratio was Braw/BpacketB_{\mathrm{raw}} / B_{\mathrm{packet}}, and savings relative to a comparator were 1−Bproposed/Bcomparator1 - B_{\mathrm{proposed}} / B_{\mathrm{comparator}}. Packet sizes included all headers, payloads and required replacement data.

Reference sizes used 4 bytes per scalar voxel, six fields for iDTT and two for QAT. Throughput was uncompressed bytes divided by elapsed time, reported in MiB/s, where 1 MiB=2201\ \mathrm{MiB}=2^{20} bytes.

Independent parsers and decoders checked complete packet lengths, decoded arrays and every voxel against the native reference. We measured maximum absolute error, root-mean-square error, signed mean error and violation counts. The two precision-sweep failures were independently reproduced. Verification used a separately constructed radial-band map but shared FFT and compression libraries with the encoder.

Native and decoded images used the same locations and display limits. Full, unclipped volumes were used for numerical checks. The extra organoid illustration in Figure 2 was not part of its 16-volume benchmark. Data images were not sharpened, denoised or generatively modified. Timing repetitions measured computational variability, not biological replication.

Numerical Reporting and Error-bound Decisions

Results are generally reported to three significant figures. Four are used when needed to show a small error-limit violation. Compliance was decided from unrounded decoded differences; a value that rounds to the limit can still lie below it. For example, the iDTT deviatoric maximum in Figure 5 rounds to 3.00×10−53.00 \times10^{-5} but was below that limit. Savings and speedups were also calculated before rounding. Counts, array dimensions and identifiers remain exact; full-precision results and settings are retained in the benchmark records.

Reproduction Materials and Software Checks

The reproduction materials contain exact input lists and hashes, fixed codec settings, software versions, benchmark scripts and independent decoders. The HT configuration uses float32 working arrays, native float64 references, an orthonormal real FFT, four radial bands, ΔF=5ε\Delta_{F}=5\varepsilon, Pcodec level 6 and residual spacing 1.98ε1.98\varepsilon. Packet-format records specify the header and payload lengths used for complete-size accounting.

The corrected float-RI path in HTPack v0.6.0b1 reproduced the compressed coefficient/residual payloads and decoded arrays for all 32 benchmark inputs, with zero nativereference violations. Its smaller headers changed packet size slightly. Independent decoding confirmed the results. The encoder rejects the two smaller-tolerance failures in Figure 4. These checks do not cover the separate 1,024-chunk cohort used in the earlier coauthor review. The distributed HTPack encoder includes additional certification and checksum operations; its timings are not interchangeable with the research-code timings defined in S6.

FixedScaleOffset-Pcodec Storage Control

We tested FixedScaleOffset-Pcodec on the same 32 inputs as Figure 2, using grid spacings 10−410^{-4} and 1.98×10−41.98 \times10^{-4} and Pcodec levels 6 and 12. Packets included filter parameters, shape and framing. All 128 packets met the native-reference 10−410^{-4} error limit without voxel replacement. Runtime and memory were not measured.

At the coarser spacing, compression ratios at levels 6/12 were 7.15/7.27 for HeLa and 8.15/8.33 for organoid. At level 12, Fourier packets were 7.81% and 12.3% smaller, respectively. Maximum errors were approximately 9.91×10−59.91 \times10^{-5}. The finer grid gave approximately half the assigned error limit; it was not a comparison at matched measured maximum error.

Coefficient and Residual Distributions

We analyzed the same 16 HeLa and 16 organoid inputs used in Figure 2 at the primary tolerance, with the same Fourier coefficient spacing, residual spacing and Pcodec level 6. The complete Fourier packets matched the original packets by SHA-256. Zero-valued quantized real and imaginary Fourier scalars accounted for 55.7% of HeLa and 70.5% of organoid coefficients. Residual integers of −1-1, 0 or +1+1 accounted for 89.2% and 92.7%, respectively. These fractions describe the tested representation; zero-valued coefficients were not identified as measurement noise.

Across the 16 inputs in each cohort, Fourier coefficient payloads totaled 8.47 MB for HeLa and 6.17 MB for organoids, while residual payloads totaled 8.52 and 7.93 MB (MB = 10610^{6} bytes). Including headers, complete Fourier packets totaled 17.0 and 14.1 MB, compared with 18.8 and 16.5 MB for direct spatial quantization at the same residual grid spacing and Pcodec level. The net storage reductions were 9.28% and 14.2%, after paying for the Fourier coefficients and additional metadata.

To examine grouping, we encoded exactly the same coefficient integers as two real/imaginary streams in FFT order instead of eight band/component streams. Coefficient payloads increased to 9.69 and 6.92 MB. Holding the residual and existing header fixed, this payload substitution reduced the net savings to 2.76% and 9.66%. This comparison changes grouping and ordering together; it is not a universal causal attribution or a measurement of a fully implemented two-stream packet format. The measured benefit depends on the inputs, quantization and entropy coder.

References

  1. [1]G. Kim et al., “Holotomography,” Nat. Rev. Methods Primers 4(1), 51 (2024) [doi:10.1038/s43586-024-00327-1].DOI
  2. [2]Y. Park, C. Depeursinge, and G. Popescu, “Quantitative phase imaging in biomedicine,” Nat. Photonics 12(10), 578–589 (2018) [doi:10.1038/s41566-018-0253-x].DOI
  3. [3]H. Hugonnet et al., “Vectorial inverse scattering for dielectric tensor tomography: overcoming challenges of reconstruction of highly scattering birefringent samples,” Opt. Express 31(18), 29654–29663 (2023) [doi:10.1364/OE.494773].DOI
  4. [4]J. Park et al., “Revealing 3D microanatomical structures of unlabeled thick cancer tissues using holotomography and virtual H&E staining,” Nat. Commun. 16(1), 4781 (2025) [doi:10.1038/s41467-025-59820-0].DOI
  5. [5]C. Oh et al., “Digital aberration correction for enhanced thick tissue imaging exploiting aberration matrix and tilt-tilt correlation from the optical memory effect,” Nat. Commun. 16(1), 1685 (2025) [doi:10.1038/s41467-025-56865-z].DOI
  6. [6]H. Hugonnet et al., “Multiscale label-free volumetric holographic histopathology of thick-tissue slides with subcellular resolution,” Adv. Photon. 3(2), 026004 (2021) [doi:10.1117/1.AP.3.2.026004].DOI
  7. [7]M. J. Lee et al., “Long-term three-dimensional high-resolution imaging of live unlabeled small intestinal organoids via low-coherence holotomography,” Exp. Mol. Med. 56(10), 2162–2170 (2024) [doi:10.1038/s12276-024-01312-0].DOI
  8. [8]J. Cho et al., “Label-free 4D holotomography with depth-adaptive segmentation for quantitative analysis of lipid droplet dynamics in hepatic organoids,” Biophotonics Discuss. 3(3), 1–18, SPIE-Intl Soc Optical Eng (2026) [doi:10.1117/1.BIOS.3.3.035005].DOI
  9. [9]P. T. Brown et al., “Fourier synthesis optical diffraction tomography for kilohertz rate volumetric imaging,” Sci. Adv. 11(33), eadr8004 (2025) [doi:10.1126/sciadv.adr8004].DOI
  10. [10]J. Hong et al., “High-speed holotomography of live cells and tissues using multi-pattern sparse axial scanning,” Opt. Express 33(22), 45708 (2025) [doi:10.1364/OE.577207].DOI
  11. [11]J. Moore et al., “OME-NGFF: a next-generation file format for expanding bioimaging data-access strategies,” Nat. Methods 18(12), 1496–1498 (2021) [doi:10.1038/s41592-021-01326-w].DOI
  12. [12]J. Moore et al., “OME-Zarr: a cloud-optimized bioimaging file format with international community support,” Histochemistry and Cell Biology 160(3), 223–251 (2023) [doi:10.1007/s00418-023-02209-1].DOI
  13. [13]D. Lee et al., “Compression benchmarking of holotomography data using OME-Zarr format,” PLOS ONE 21(7), e0351560 (2026) [doi:10.1371/journal.pone.0351560].DOI
  14. [14]C. S. Zender, “Bit Grooming: statistically accurate precision-preserving quantization with compression, evaluated in the netCDF Operators (NCO, v4.4.8+),” Geosci. Model Dev. 9(9), 3199–3211 (2016) [doi:10.5194/gmd-9-3199-2016].DOI
  15. [15]M. Klöwer et al., “Compressing atmospheric data into its real information content,” Nature Computational Science 1(11), 713–724 (2021) [doi:10.1038/s43588-021-00156-2].DOI
  16. [16]M. Loncaric, N. Jeppesen, and B. Zinberg, “Pcodec: Better Compression for Numerical Sequences,” 2502.06112, arXiv (2025) [doi:10.48550/arXiv.2502.06112].DOI
  17. [17]X. Delaunay, A. Courtois, and F. Gouillon, “Evaluation of lossless and lossy algorithms for the compression of scientific datasets in netCDF-4 or HDF5 files,” Geoscientific Model Development 12(9), 4099–4113 (2019) [doi:10.5194/gmd-12-4099-2019].DOI
  18. [18]M. Ziemczonok, A. Kuś, and M. Kujawińska, “Optical diffraction tomography meets metrology — Measurement accuracy on cellular and subcellular level,” Measurement 195, 111106 (2022) [doi:10.1016/j.measurement.2022.111106].DOI
  19. [19]A. C. Fluty and S. J. Ludtke, “Precision requirements and data compression in CryoEM/CryoET,” Journal of Structural Biology 214(3), 107875 (2022) [doi:10.1016/j.jsb.2022.107875].DOI
  20. [20]S. Di et al., “A Survey on Error-Bounded Lossy Compression for Scientific Datasets,” ACM Computing Surveys 57(11), 1–38 (2025) [doi:10.1145/3733104].DOI
  21. [21]S. Di and F. Cappello, “Fast Error-Bounded Lossy HPC Data Compression with SZ,” in 2016 IEEE International Parallel and Distributed Processing Symposium (IPDPS), pp. 730–739 (2016) [doi:10.1109/ipdps.2016.11].DOI
  22. [22]D. Tao et al., “Significantly Improving Lossy Compression for Scientific Data Sets Based on Multidimensional Prediction and Error-Controlled Quantization,” in 2017 IEEE International Parallel and Distributed Processing Symposium (IPDPS), pp. 1129–1139 (2017) [doi:10.1109/ipdps.2017.115].DOI
  23. [23]P. Lindstrom, “Fixed-Rate Compressed Floating-Point Arrays,” IEEE Transactions on Visualization and Computer Graphics 20(12), 2674–2683 (2014) [doi:10.1109/tvcg.2014.2346458].DOI
  24. [24]R. Ballester-Ripoll, P. Lindstrom, and R. Pajarola, “TTHRESH: Tensor Compression for Multidimensional Visual Data,” IEEE Transactions on Visualization and Computer Graphics 26(9), 2891–2903 (2020) [doi:10.1109/tvcg.2019.2904063].
  25. [25]M. Ainsworth et al., “Multilevel Techniques for Compression and Reduction of Scientific Data---The Multivariate Case,” SIAM Journal on Scientific Computing 41(2), A1278–A1303 (2019) [doi:10.1137/18m1166651].DOI
  26. [26]S. Li, P. Lindstrom, and J. Clyne, “Lossy Scientific Data Compression With SPERR,” in 2023 IEEE International Parallel and Distributed Processing Symposium (IPDPS), pp. 1007–1017 (2023) [doi:10.1109/ipdps54959.2023.00104].DOI
  27. [27]C. Ren et al., “FFCZ: Fast Fourier Correction for Spectrum-Preserving Lossy Compression of Scientific Data,” Preprint 2601.01596, arXiv (2026) [doi:10.48550/arxiv.2601.01596].DOI
  28. [28]J. Diffenderfer et al., “Error Analysis of ZFP Compression for Floating-Point Data,” SIAM Journal on Scientific Computing 41(3), A1867–A1898 (2019) [doi:10.1137/18M1168832].DOI
  29. [29]J. Lee et al., “Incoherent dielectric tensor tomography for quantitative three-dimensional measurement of biaxial anisotropy,” Nat. Photonics 20(7), 773–782 (2026) [doi:10.1038/s41566-026-01897-0].DOI
  30. [30]Y. Chung et al., “Quantitative three-dimensional absorption imaging in standard brightfield microscopes,” arXiv (2026) [doi:10.48550/ARXIV.2601.15925].DOI
  31. [31]E. Wolf, “Three-dimensional structure determination of semi-transparent objects from holographic data,” Opt. Commun. 1(4), 153–156 (1969) [doi:10.1016/0030-4018(69)90052-2].DOI
  32. [32]A. J. Devaney, “A filtered backpropagation algorithm for diffraction tomography,” Ultrason. Imaging 4(4), 336–350 (1982) [doi:10.1177/016173468200400404].DOI
  33. [33]T. Kim et al., “White-light diffraction tomography of unlabelled live cells,” Nature Photon 8(3), 256–263 (2014) [doi:10.1038/nphoton.2013.350].DOI
  34. [34]H. Hugonnet, M. Lee, and Y. Park, “Optimizing illumination in three-dimensional deconvolution microscopy for accurate refractive index tomography,” Opt. Express 29(5), 6293–6301 (2021) [doi:10.1364/OE.412510].DOI
  35. [35]H. Hugonnet, M. J. Lee, and Y. K. Park, “Quantitative phase and refractive index imaging of 3D objects via optical transfer function reshaping,” Opt. Express 30(8), 13802–13809 (2022) [doi:10.1364/OE.454533].DOI
  36. [36]K. Kim et al., “High-resolution three-dimensional imaging of red blood cells parasitized by Plasmodium falciparum and in situ hemozoin crystals using optical diffraction tomography,” J. Biomed. Opt. 19(01), 1 (2013) [doi:10.1117/1.JBO.19.1.011005].DOI
  37. [37]S. Chowdhury et al., “High-resolution 3D refractive index microscopy of multiple-scattering samples from intensity images,” Optica 6(9), 1211 (2019) [doi:10.1364/OPTICA.6.001211].DOI
  38. [39]P. Lindstrom et al., “ZFP: A compressed array representation for numerical computations,” The International Journal of High Performance Computing Applications 39(1), 104–122 (2025) [doi:10.1177/10943420241284023].DOI
  39. [40]P. Schelkens et al., “Compression strategies for digital holograms in biomedical and multimedia applications,” Light Adv. Manuf. 3(3), 601–621 (2022) [doi:10.37188/lam.2022.040].DOI
  40. [41]T. J. Naughton et al., “Compression of digital holograms for three-dimensional object reconstruction and recognition,” Applied Optics 41(20), 4124 (2002) [doi:10.1364/ao.41.004124].DOI
  41. [42]A. E. Shortt, T. J. Naughton, and B. Javidi, “A companding approach for nonuniform quantization of digital holograms of three-dimensional objects,” Optics Express 14(12), 5129 (2006) [doi:10.1364/oe.14.005129].DOI
  42. [43]P. Stępień et al., “Spatial bandwidth-optimized compression of image plane off-axis holograms with image and video codecs,” Optics Express 28(19), 27873 (2020) [doi:10.1364/oe.398598].DOI
  43. [44]R. K. Muhamad et al., “Off-axis image plane hologram compression in holographic tomography – metrological assessment,” Optics Express 30(3), 4261 (2022) [doi:10.1364/oe.449932].DOI
  44. [45]R. Kizhakkumkara Muhamad et al., “INTERFERE, short-time Fourier-transform-based compression of complex-valued holograms with bit depth and range-adaptive quantization,” Applied Optics 63(25), 6740 (2024) [doi:10.1364/ao.531422].DOI
  45. [46]P. Memmolo et al., “Loss Minimized Data Reduction in Single-Cell Tomographic Phase Microscopy Using 3D Zernike Descriptors,” Intel Comput 2, 0010 (2023) [doi:10.34133/icomputing.0010].DOI
  46. [47]T. Chiueh et al., “Integrated volume compression and visualization,” in Proceedings. Visualization ’97 (Cat. No. 97CB36155), pp. 329-336, (1997) [doi:10.1109/visual.1997.663900].
  47. [48]M. C. Restivo et al., “A framework for constraining image SNR loss due to MR raw data compression,” Magnetic Resonance Materials in Physics, Biology and Medicine 32(2), 213–225 (2019) [doi:10.1007/s10334-018-0709-5].DOI
  48. [49]C. Park, S. Shin, and Y. Park, “Generalized quantification of three-dimensional resolution in optical diffraction tomography using the projection of maximal spatial bandwidths,” J. Opt. Soc. Am. A 35(11), 1891–1898 (2018) [doi:10.1364/josaa.35.001891].DOI
  49. [53]M. Ainsworth et al., “Multilevel Techniques for Compression and Reduction of Scientific Data-Quantitative Control of Accuracy in Derived Quantities,” SIAM Journal on Scientific Computing 41(4), A2146–A2171 (2019) [doi:10.1137/18m1208885].DOI
  50. [54]S. Aknoun et al., “Living cell dry mass measurement using quantitative phase imaging with quadriwave lateral shearing interferometry: an accuracy and sensitivity discussion,” J. Biomed. Opt 20(12), 126009 (2015) [doi:10.1117/1.JBO.20.12.126009].DOI
  51. [55]C. Oh et al., “Morphology-preserving holotomography: quantitative analysis of 3D organoid dynamics,” IEEE Trans. Med. Imaging 45(8), 4380–4394 (2026) [doi:10.1109/TMI.2026.3698497].DOI
  52. [56]X. Liu et al., “Computationally enhanced quantitative phase microscopy reveals autonomous oscillations in mammalian cell growth,” Proceedings of the National Academy of Sciences 117(44), 27388–27399, Proceedings of the National Academy of Sciences (2020) [doi:10.1073/pnas.2002152117].DOI
  53. [57]K. Kim et al., “Three-dimensional label-free imaging and quantification of lipid droplets in live hepatocytes,” Sci Rep 6(1), 36815 (2016) [doi:10.1038/srep36815].DOI

Paper details

vlnc:W_cachndxx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Codex

View statement
Looking for related papers…
Contents