betaFeedback

WorksEngineering

Hybrid quantum learning for imbalanced blood-cell classification: domain adaptation and classifier design

Abstract

Limited annotated data, class imbalance, and morphological similarity make white blood cell classification challenging. We benchmark hybrid variational quantum classifier heads against an affine baseline on 821 microscopy images across seven morphological classes. Using ImageNet-initialised ResNet34 backbones and repeated stratified five-fold cross-validation, we first compare direct transfer with intermediate adaptation on MLL23, a larger peripheral blood cell microscopy dataset. To guide the head comparison, we explore different encoding strategies, variational depth, and the number of independent circuit blocks. We then compare the classifier performance using the affine head and two hybrid heads across multiple random seeds, under a shared training and evaluation protocol. For this comparison, all backbones share the same MLL23-adapted initialisation. Intermediate adaptation increases mean macro-F1 across all four affine-head fine-tuning configurations. The largest gain occurs with a frozen backbone, where macro-F1 rises from 0.6850.685 to 0.7860.786, a relative improvement of approximately 14.7%14.7\%. The shallow two-block hybrid head achieves a mean pooled out-of-fold macro-F1 of 0.865±0.0100.865 \pm 0.010, compared with 0.879±0.0080.879 \pm 0.008 for the affine head. The corresponding balanced accuracies are 0.886±0.0080.886 \pm 0.008 and 0.889±0.0100.889 \pm 0.010, respectively. In the architectural exploration, the two-block hybrid head achieves the highest macro-F1 and balanced accuracy among the evaluated hybrid configurations, while the effect of encoding varies across the considered metrics. Our results help guide the design and evaluation of hybrid classifier heads for imbalanced microscopy by linking architectural choices to overall performance and class-specific precision-recall trade-offs.

*Corresponding author: Gabriele D’Angeli. Email: [email protected], [email protected].

Introduction

Image classification is challenging when labelled examples are scarce, classes are unevenly represented, and visual differences between categories are subtle. White blood cell (WBC) classification provides a concrete example: morphologically similar categories can have very different numbers of available training samples. Recent research in haematological image analysis has examined class imbalance as an important factor affecting recognition across cell categories [31]. In this setting, high overall classification accuracy can coexist with poor performance on under-represented classes. Meaningful evaluation therefore requires both aggregate metrics and class-specific analysis.

Transfer learning helps address the scarcity of labelled images by reusing representations learnt from larger datasets. In haematological microscopy, Koch et al. [13] showed that models trained on cell images learn features that generalise across datasets. Class imbalance can also be addressed during training through class-weighted losses, resampling according to class frequency, and more frequent augmentation of under-represented classes. In these pipelines, the backbone extracts image features and the classifier head maps them to class predictions. Comparing heads can therefore reveal whether alternative ways of processing these features improve recognition of under-represented cell categories.

Quantum machine learning (QML) offers alternative approaches to classifier design. One motivation for studying quantum models with limited training data comes from theoretical work on generalisation. Caro et al. [3] established upper bounds on the generalisation error that depend on the size of the training set and the number of trainable circuit gates. The performance of these models relative to classical alternatives remains an empirical question. Hybrid quantum-classical transfer learning provides a practical framework for such comparisons in image classification. In the approach of Mari et al. [16], a pretrained classical network extracts image features, which are then processed by a variational quantum circuit (VQC) and a classical readout.

Previous studies have explored quantum and hybrid models in medical imaging, with some reporting improvements over classical models and others reporting comparable or lower performance. Matic et al. [17] reported performance comparable to that of classical convolutional neural networks for breast ultrasound and abdominal CT images in simulation. Similar results were obtained for chest radiographs and retinal images in simulation, although performance declined in some configurations when both training and inference were performed on quantum hardware [14]. Hybrid models underperformed their classical counterparts in imbalanced multilabel chest X-ray classification [5], while improvements in macro-F1 were observed in blood-cell microscopy, including with smaller training sets [6].

A systematic review of QML for digital health found no consistent performance advantage over classical methods and highlighted the need for more rigorous comparisons [10]. Such comparisons need to account for both the choice of classical baselines and the role of trainable classical components, since gains cannot automatically be attributed to the quantum circuit [2, 30]. These considerations motivate evaluating hybrid heads in the context of the complete learning pipeline.

We use MicroscopyRGB, a dataset of 821 peripheral blood cell images derived from Raw-Microscopy [19 b]. We compare direct ImageNet transfer with intermediate adaptation on MLL23, a larger peripheral blood cell microscopy dataset. We also evaluate affine and hybrid classifier heads under a shared training and evaluation protocol. An exploration of encoding and circuit architecture guides the head comparison. Whereas Cruz et al. [6] compared intermediate transformations, we evaluate the replacement of the complete classifier head.

This study therefore addresses three research questions:

  • Q1 Does intermediate adaptation on MLL23 improve target-task performance compared with direct ImageNet transfer, and how does its effect depend on the extent of backbone fine-tuning?

  • Q2 How do hybrid VQC heads compare with an affine head in terms of aggregate performance and recognition of under-represented classes under matched training and evaluation conditions?

  • Q3 How do encoding strategy, the number of variational layers, and the number of independent circuit blocks relate to aggregate performance and class-specific precision-recall trade-offs?

This study jointly assesses domain adaptation and classifier-head design under a shared protocol for microscopy image classification, considering both aggregate performance and class-specific errors.

Methods

The experimental workflow comprises three stages. First, we compare direct ImageNet transfer and intermediate MLL23 adaptation across four backbone fine-tuning configurations using affine heads and three random seeds. The MLL23-adapted initialisation with fine-tuning of layers 3-4 is selected for the subsequent classifier-head comparison. Second, the seven hybrid configurations are explored using one seed under the same target-task training and evaluation protocol. Based on aggregate performance in this exploration, we select the two-block hybrid head for repeated evaluation. In the third stage, this model and a shallow single-block reference are evaluated using two additional seeds. Both configurations use 12 qubits and one variational layer per block. The final comparison therefore includes the affine baseline and these two hybrid heads, each evaluated across three seeds.

Data and transfer-learning setup

Target dataset

The Raw-Microscopy dataset contains WBC images acquired in our laboratory from peripheral blood smear slides and labelled by a trained haematological cytologist, as described in the original dataset publication [19 b]. For this study, we reuse a single pre-generated RGB representation with 256×256256 \times256-pixel images, hereafter referred to as MicroscopyRGB.

Of the original 940 images, 30 labelled as unclassified (UNC) are excluded because they do not represent a defined target category. The eight band-neutrophil images (NGB) are also excluded because this class provides very limited support for training and stratified evaluation. We then manually inspect the dataset image by image and remove 81 duplicate images, leaving a final dataset of 821 images across seven morphological categories: basophils, eosinophils, smudge cells, atypical lymphocytes, typical lymphocytes, monocytes, and segmented neutrophils (Table 1).

Class labelCell typeSamples
BASBasophil18
EOSEosinophil43
KSCSmudge cell158
LYAAtypical lymphocyte30
LYTTypical lymphocyte224
MONMonocyte180
NGSSegmented neutrophil168
Total821

Table 1. Class distribution of the final MicroscopyRGB dataset (N=821N = 821).

Images are loaded as 256×256×3256 \times256 \times3 TIFF files and rescaled to [0,1][0, 1]. Each colour channel is then standardised using the mean pixel intensity and standard deviation for that channel: μ=(0.5566,0.4607,0.5923)\mu= (0.5566, 0.4607, 0.5923) and σ=(0.0983,0.1417,0.0667)\sigma= (0.0983, 0.1417, 0.0667).

Transfer-learning strategies

The MLL23 dataset contains 41,621 images of peripheral blood cells across 18 morphological classes [29, 28]. Images are acquired by automated microscopy at the Munich Leukemia Laboratory and released as 288×288288 \times288-pixel TIFF files. Each image is annotated by one of five cytomorphology experts. Several morphological categories overlap with those in MicroscopyRGB, motivating its use for intermediate adaptation.

We compare two transfer-learning paths, both starting from an ImageNet-pretrained ResNet34 [9, 24]. Direct transfer uses the ImageNet weights to initialise the backbone for training on MicroscopyRGB. The second path includes intermediate adaptation on MLL23 before training on MicroscopyRGB.

For intermediate adaptation, the ImageNet-pretrained ResNet34 is fine-tuned on MLL23 using a stratified 80/20 training-validation split. Training uses focal loss with γ=0.5\gamma= 0.5, label smoothing of 0.05, and effective-number class weighting. The checkpoint with the highest validation macro-F1 is selected. Its backbone weights are then used to initialise the target-task model, which is trained on MicroscopyRGB with a newly initialised classifier head.

Affine and hybrid classification pipelines

Feature extraction and affine baseline

Both pipelines use ResNet34 to extract a compact image representation [11]. After the final residual stage and global average pooling, the backbone maps an input image I\mathbf{I} to the 512-dimensional feature vector

h=fϕ(I)∈R512,(1)\mathbf{h} = f_{\phi}(\mathbf{I}) \in\mathbb{R}^{512}, \tag*{(1)}

where ϕ\phi denotes the parameters of the ResNet34 backbone.

The transfer-learning experiments use either ImageNet or MLL23-adapted initialisation. For the affine-hybrid comparison, all backbones start from the same MLL23-adapted weights. Residual layers 3-4 are fine-tuned jointly with each head, while the stem and residual layers 1-2 remain frozen. This allows the backbone features to adapt to each classifier head.

Figure 1 provides an overview of the affine and hybrid architectures used for the head comparison.

Affine and hybrid classification pipelines used for the head comparison

Figure 1. Affine and hybrid classification pipelines used for the head comparison. Both ResNet34 backbones start from the same MLL23-adapted checkpoint and are fine-tuned separately. The hybrid head uses BB independent 12-qubit blocks, each with LL variational layers. The projection dimension is D=36BD = 36B for standard re-uploading and D=36BLD = 36BL for incremental re-uploading. Concatenation yields 36B36B expectation values for the classical readout. Input preprocessing and the tanh bounding in Eq. (3) are omitted

The affine head replaces the original ImageNet classification layer with a newly initialised affine map from the backbone features to the seven target logits, defined by

zC=WCh+bC,(2)\mathbf{z}_{\mathrm{C}} = W_{\mathrm{C}}\mathbf{h} + \mathbf{b}_{\mathrm{C}}, \tag*{(2)}

where WC∈R7×512W_{\mathrm{C}} \in\mathbb{R}^{7 \times512} is the weight matrix, bC∈R7\mathbf{b}_{\mathrm{C}} \in\mathbb{R}^{7} is the trainable bias vector, and zC∈R7\mathbf{z}_{\mathrm{C}} \in\mathbb{R}^{7} contains one logit for each class.

Hybrid head and input encoding

The hybrid head comprises a classical projection to bounded circuit inputs, one or more quantum circuit blocks, and a classical readout. Encoding influences the functions that the circuit can represent. Schuld et al. [25] characterise quantum models with common data-encoding gates through a Fourier representation, showing how encoding operations determine the accessible frequencies and how repeated encoding can enlarge this spectrum. Data re-uploading alternates input-dependent gates with trainable transformations [22]. These studies motivate the comparison of encoding strategies described below.

We compare standard re-uploading, which reuses the same projected features at successive layers, with incremental re-uploading, which supplies a different feature slice to each layer. The latter follows the principle introduced by Periyasamy et al. [23], adapted here to features learnt by a classical backbone and projection rather than to a fully quantum classifier.

Let BB denote the number of independent circuit blocks, each containing qb=12q_b = 12 qubits, and let q=Bqbq = Bq_b denote their aggregate qubit count. Let LL denote the number of variational layers per block. The projection dimension is D=3qD = 3q for standard re-uploading and D=3qLD = 3qL for incremental re-uploading.

The ResNet34 feature vector is layer-normalised, projected by a trainable affine map, and transformed element-wise using a scaled tanh:

x=3π4tanh⁡[WALN⁡(h)+bA],(3)\mathbf{x} = \frac{3\pi}{4}\tanh\left[\mathbf{W}_A\operatorname{LN}(\mathbf{h})+\mathbf{b}_A\right], \tag*{(3)}

where LN⁡\operatorname{LN} denotes layer normalisation. The matrix WA∈RD×512\mathbf{W}_A \in\mathbb{R}^{D \times512} and the vector bA∈RD\mathbf{b}_A \in\mathbb{R}^{D} are the trainable projection weights and bias, and the subscript A denotes the angle-input projection. The scaled tanh bounds each component of x\mathbf{x} to the interval (−3π/4,3π/4)(-3\pi/4,3\pi/4).

With standard re-uploading, the same 3q3q-dimensional input vector is supplied to every layer:

ξl=x,l=1,…,L.(4)\boldsymbol{\xi}_l = \mathbf{x}, \qquad l = 1,\ldots,L. \tag*{(4)}

With incremental re-uploading, the 3qL3qL-dimensional projection is split into LL layer-specific input vectors, each of dimension 3q3q:

x=[ξ1T ⋯ ξLT]T,ξl∈R3q.(5)\mathbf{x} = \left[\boldsymbol{\xi}_1^{\mathsf{T}}\ \cdots\ \boldsymbol{\xi}_L^{\mathsf{T}}\right]^{\mathsf{T}}, \qquad\boldsymbol{\xi}_l \in\mathbb{R}^{3q}. \tag*{(5)}

For B>1B>1, each ξl\boldsymbol{\xi}_l is further partitioned into BB contiguous block-specific vectors ξl(b)∈R3qb\boldsymbol{\xi}_l^{(b)} \in\mathbb{R}^{3qb}. We denote by ξlia,(b)\xi_{li}^{a,(b)} the component assigned to rotation axis a∈{X,Y,Z}a \in\{X,Y,Z\} of qubit ii in block bb at layer ll.

Circuit structure and readout

Multiple circuits of limited width have been used in hybrid image-classification architectures [27]. In our implementation, each block contains 12 qubits and is simulated independently, with no quantum operations connecting different blocks. This design allows the number of encoded features and readout values to increase while keeping the width of each simulated circuit fixed. All trainable components are optimised jointly within each hybrid pipeline.

The joint state of the qubits in block bb is initialised as

∣ψ0(b)⟩=∣0⟩⊗qb.(6)\lvert\psi_0^{(b)}\rangle= \lvert0\rangle^{\otimes q_b}. \tag*{(6)}

At circuit layer ll, the rotation angle associated with axis a∈{X,Y,Z}a \in\{X,Y,Z\}, qubit ii, and block bb is defined as

γlia,(b)=θlia,(b)+ωlia,(b)ξlia,(b),a∈{X,Y,Z},(7)\gamma_{li}^{a,(b)} = \theta_{li}^{a,(b)} + \omega_{li}^{a,(b)}\xi_{li}^{a,(b)}, \qquad a \in\{X,Y,Z\}, \tag*{(7)}

where θlia,(b)\theta_{li}^{a,(b)} is a trainable rotation offset and ωlia,(b)\omega_{li}^{a,(b)} is a trainable encoding scale.

The local single-qubit transformations at layer ll are collected into the operator

Urot(l,b)=⨂i=1qb[RY(γliY,(b))RZ(γliZ,(b))RX(γliX,(b))].(8)U_{\mathrm{rot}}^{(l,b)} = \bigotimes_{i=1}^{q_b}\left[R_Y\left(\gamma_{li}^{Y,(b)}\right)R_Z\left(\gamma_{li}^{Z,(b)}\right)R_X\left(\gamma_{li}^{X,(b)}\right)\right]. \tag*{(8)}

With the standard right-to-left convention for operator composition, RXR_X, RZR_Z, and RYR_Y are therefore applied sequentially to each qubit.

Each variational layer applies these local rotations followed by a nearest-neighbour ring of trainable Ising-ZZ interactions, using a hardware-efficient design [12]:

UZZ(l,b)=∏(i,j)∈Ebexp⁡(−i2φij(l,b)ZiZj),(9)U_{\mathrm{ZZ}}^{(l,b)} = \prod_{(i,j)\in\mathcal{E}_b}\exp\left(-\frac{\mathrm{i}}{2}\varphi_{ij}^{(l,b)}Z_iZ_j\right), \tag*{(9)}

where φij(l,b)\varphi_{ij}^{(l,b)} is the trainable Ising-ZZ interaction angle for edge (i,j)(i,j) in block bb at layer ll. The edge set Eb\mathcal{E}_b contains the nearest-neighbour edges of the bthb^{\mathrm{th}} block, including the closing edge connecting the last qubit to the first.

The state of block bb therefore evolves according to

∣ψl(b)⟩=UZZ(l,b)Urot(l,b)∣ψl−1(b)⟩,l=1,…,L.(10)\lvert\psi_l^{(b)}\rangle= U_{\mathrm{ZZ}}^{(l,b)}U_{\mathrm{rot}}^{(l,b)}\lvert\psi_{l-1}^{(b)}\rangle, \qquad l = 1,\ldots,L. \tag*{(10)}

After the final layer LL, a trainable Rot⁡(αi(b),βi(b),δi(b))\operatorname{Rot}(\alpha_i^{(b)},\beta_i^{(b)},\delta_i^{(b)}) gate is applied to each qubit before readout:

Umeas(b)=⨂i=1qbRot⁡(αi(b),βi(b),δi(b)),∣ψ~L(b)⟩=Umeas(b)∣ψL(b)⟩(11)U_{\mathrm{meas}}^{(b)}=\bigotimes_{i=1}^{q_b}\operatorname{Rot}(\alpha_i^{(b)},\beta_i^{(b)},\delta_i^{(b)}),\qquad|\widetilde{\psi}_L^{(b)}\rangle=U_{\mathrm{meas}}^{(b)}|\psi_L^{(b)}\rangle \tag*{(11)}

The local Bloch-vector readout of block bb, evaluated on ∣ψ~L(b)⟩|\widetilde{\psi}_L^{(b)}\rangle, is

r(b)=(⟨X1⟩,…,⟨Xqb⟩, ⟨Y1⟩,…,⟨Yqb⟩, ⟨Z1⟩,…,⟨Zqb⟩)∈R3qb.(12)\mathbf{r}^{(b)}=\left(\langle X_1\rangle,\ldots,\langle X_{q_b}\rangle,\ \langle Y_1\rangle,\ldots,\langle Y_{q_b}\rangle,\ \langle Z_1\rangle,\ldots,\langle Z_{q_b}\rangle\right)\in\mathbb{R}^{3q_b}. \tag*{(12)}

Figure 2 summarises the variational layers and Bloch-vector readout of one 12-qubit block.

Figure 2. Structure of one 12-qubit VQC block bb. (a) Each of the LL variational layers applies RXR_X, RZR_Z, and RYR_Y rotations sequentially, with trainable angle offsets and encoding scales, followed by a trainable Ising-ZZZZ ring. The final Rot⁡i=Rot⁡(αi,βi,δi)\operatorname{Rot}_i=\operatorname{Rot}(\alpha_i,\beta_i,\delta_i) gates are applied once, after layer LL. Expectation values of XX, YY, and ZZ on each qubit form a 36-component readout vector. (b) Nearest-neighbour ring topology, including the 12-1 edge. The block index bb is suppressed in local gate labels

For multiple blocks, these vectors are concatenated in block order. The resulting vector r∈R3q\mathbf{r}\in\mathbb{R}^{3q} is batch-normalised and mapped to seven logits:

zH=WHBN⁡(r)+bH,(13)\mathbf{z}_{H}=W_{H}\operatorname{BN}(\mathbf{r})+\mathbf{b}_{H}, \tag*{(13)}

where BN⁡\operatorname{BN} denotes batch normalisation, and the subscript H denotes the hybrid head. The readout has trainable weights WH∈R7×3qW_H\in\mathbb{R}^{7\times3q} and bias bH∈R7\mathbf{b}_H\in\mathbb{R}^{7}. Its output zH∈R7\mathbf{z}_H\in\mathbb{R}^{7} contains one logit for each target class.

Initialisation, parameter counts, and comparison scope

Encoding scales are initialised to one. Rotation offsets, Ising-ZZZZ angles, and final measurement-basis rotation parameters are independently drawn from N(0,0.01)\mathcal{N}(0,0.01). Circuit expectation values are computed analytically using noiseless statevector simulation.

Each block is a 12-qubit ring with ∣Eb∣=qb\lvert E_b\rvert=q_b edges. The number of trainable circuit parameters is therefore

Ncircuit=B(6Lqb+L∣Eb∣+3qb)=Bqb(7L+3)=q(7L+3).(14)N_{\mathrm{circuit}}=B\left(6Lq_b+L\lvert E_b\rvert+3q_b\right) =Bq_b(7L+3) =q(7L+3). \tag*{(14)}

The first term contains the rotation offsets and encoding scales, the second the Ising-ZZZZ angles, and the third the final measurement-basis rotations.

Table 2 reports parameter counts for the shallow configurations. The hybrid heads contain more trainable post-backbone parameters than the affine baseline, although the difference is small relative to the backbone parameter count. The comparison evaluates complete head replacement and therefore does not isolate the contribution of the VQC from that of the additional classical projection, normalisation, and readout.

ModelMax. circuit widthPost-backbone headCircuit params.TotalTrainable
ResNet34 + affine head-3,591-21,288,26319,940,359
Hybrid B1/L1B1/L11219,94312021,304,61519,956,711
Hybrid B2/L1B2/L11238,85524021,323,52719,975,623
Hybrid B3/L1B3/L11257,76736021,342,43919,994,535

Table 2. Parameter counts for the shallow MLL23-adapted configurations. Hybrid rows use L=1L = 1, an affine angle projection with tanh bounding, Bloch-vector readout, and trainable Ising-ZZZZ and final measurement-basis rotations. Trainable totals assume the layer-3-4 fine-tuning configuration. Circuit parameters are included in the post-backbone head count.

Training and evaluation protocol

The target-task head comparison uses shared fold assignments, backbone initialisation and fine-tuning settings, augmentation, imbalance handling, and optimisation settings. Intermediate adaptation on MLL23 is conducted separately using the procedure described above.

Cross-validation design

Target-task performance is assessed using shuffled five-fold cross-validation stratified by class. Each image appears in a validation fold exactly once per seed. No separate test set is allocated, given the limited total sample size and low support of several classes. Performance is therefore estimated from pooled out-of-fold (OOF) predictions.

The transfer-learning and classifier-head comparisons use three randomly chosen seeds: 250, 9, and 122. The architectural exploration uses the first of these, and the two selected hybrid heads are then evaluated with the other two. All compared pipelines use identical fold assignments within each seed. The separate stratified training-validation split for intermediate MLL23 adaptation uses seed 20.

Loss and imbalance handling

For each fold, the model is trained for 50 epochs with a batch size of 64. The objective is class-weighted cross-entropy with label smoothing ϵ=0.05\epsilon= 0.05. For C=7C = 7 classes, the smoothed target distribution is

y~ic=(1−ϵ)1[c=yi]+ϵC.(15)\widetilde{y}_{ic} = (1-\epsilon)\mathbf{1}[c=y_i] + \frac{\epsilon}{C}. \tag*{(15)}

Class weights are calculated separately from the training partition of each fold using the effective number of observations [7]. For class cc with ncn_c training images,

Ec=1−βnc1−β,wc=CEc−1∑k=1CEk−1,β=0.999.(16)E_c = \frac{1-\beta^{n_c}}{1-\beta}, \qquad w_c = \frac{C E_c^{-1}}{\sum_{k=1}^{C} E_k^{-1}}, \qquad\beta= 0.999. \tag*{(16)}

The weights have unit mean across classes and are applied within the smoothed cross-entropy loss to increase the contribution of under-represented categories.

Training images are also sampled with replacement using inverse-frequency weights si=1/nyis_i = 1/n_{y_i}. Because the weights sum to one within each class, all classes have equal expected sampling probability. Each epoch contains as many draws as images in the training fold, allowing an image to be sampled more than once or omitted in a given epoch. The same loss and sampling procedures are used across target-task comparisons.

Conditional data augmentation

Augmentation is applied only to training images, before channel-wise normalisation. Eligibility depends on class support in the current training fold. Images from classes with at most 30 training observations, BAS and LYA, are always eligible for augmentation, while images from the remaining classes are eligible with probability 0.3.

Geometric transformations comprise horizontal and vertical flips, each applied with probability 0.5, and a randomly selected rotation of 90∘90^\circ, 180∘180^\circ, or 270∘270^\circ, applied with probability 0.5. Brightness and contrast are adjusted with probability 0.35. The augmentation pipeline also includes additive Gaussian noise (with probability 0.25), binomial blur (with probability 0.20), and sharpness adjustment (with probability 0.35).

Optimisation and epoch selection

All trainable components are optimised jointly using AdamW [15]. For the trainable ResNet34 layers we use a learning rate of 3×10−43 \times10^{-4}, while the model-specific post-backbone head uses 10−310^{-3}. Weight decay is set to 10−310^{-3}. In hybrid models, circuit rotation offsets, encoding scales, entangling angles, and final measurement-basis rotations use the head learning rate with zero weight decay.

A ReduceLROnPlateau scheduler monitors the validation macro-F1 score. The minimum learning rate for each parameter group is set at 1% of its initial value. For each fold, validation predictions from the epoch with the highest macro-F1 are retained for OOF aggregation.

Hybrid pipelines are trained end-to-end through the trainable backbone layers, angle projection, circuits, and classical readout. Circuits are evaluated with the lightning.gpu backend and adjoint differentiation on an NVIDIA A40 GPU.

Metrics and OOF aggregation

Macro-F1 is the primary aggregate metric. It assigns equal weight to each of the C=7C = 7 classes and is computed from the precision PcP_c and recall RcR_c of each class cc:

F1,macro=1C∑c=1C2PcRcPc+Rc,(17)F_{1,\mathrm{macro}} = \frac{1}{C} \sum_{c=1}^{C} \frac{2P_cR_c}{P_c + R_c}, \tag*{(17)}

where undefined class-specific values are set to zero.

Complementary metrics are accuracy, weighted F1, macro precision, and class-specific precision, recall, and F1. Balanced accuracy is calculated as

BA=1C∑c=1CRc.(18)\mathrm{BA} = \frac{1}{C} \sum_{c=1}^{C} R_c. \tag*{(18)}

All metrics are computed using scikit-learn [21].

For each configuration and seed, we pool validation predictions from the selected epoch of each fold, yielding one prediction for each of the 821 images. Aggregate and class-specific metrics, as well as the confusion matrix, are calculated directly from these pooled predictions.

For repeated configurations, pooled OOF metrics are calculated separately for each seed and reported as the mean and sample standard deviation across the three seeds.

Experiment tracking and computational cost

We track experiments using Weights & Biases [1] and record end-to-end training times. Hybrid runtimes are measured under noiseless statevector simulation.

Results

Results address the research questions introduced in § 1. For the transfer-learning and classifier-head comparisons, we report the mean and sample standard deviation of the evaluation metrics across three seeds. The architectural exploration uses a single seed.

Intermediate MLL23 adaptation

Intermediate MLL23 adaptation increases mean pooled OOF macro-F1 in all four affine-head configurations (Figure 3a). The largest gain occurs with a frozen backbone, where macro-F1 rises from 0.685 to 0.786, a relative improvement of approximately 14.7%. With target-task fine-tuning, the increases are smaller, ranging from 0.004 to 0.011. Aggregate metrics for all four fine-tuning configurations are reported in Table S1 in Appendix § 8.

Macro-F1 scores and confusion matrices with and without intermediate MLL23 adaptation

Figure 3. Macro-F1 scores and confusion matrices with and without intermediate MLL23 adaptation. (a) Mean pooled out-of-fold (OOF) macro-F1 for four ResNet34 fine-tuning configurations. Error bars show sample standard deviations across seeds 9, 122, and 250. (b) Confusion matrices for frozen-backbone models using direct ImageNet transfer and MLL23-adapted initialisation. For each seed, the pooled OOF matrix is normalised by true class, and corresponding entries are then averaged across the three seeds. Cell values and colours indicate mean within-class percentages. Diagonal entries represent mean class recall. All configurations use affine classifier heads and five-fold cross-validation on the same 821 images for each seed

Backbone fine-tuningInitialisationAccuracyBalanced AccuracyMacro-PrecisionMacro-F1Weighted-F1
FullImageNet0.935±0.0040.935 \pm0.0040.891±0.0090.891 \pm0.0090.861±0.0080.861 \pm0.0080.874±0.0080.874 \pm0.0080.937±0.0030.937 \pm0.003
FullMLL23-adapted0.933±0.0040.933 \pm0.0040.901±0.0120.901 \pm0.0120.869±0.0090.869 \pm0.0090.881±0.0040.881 \pm0.0040.936±0.0010.936 \pm0.001
Layers 3-4ImageNet0.932±0.0040.932 \pm0.0040.895±0.0040.895 \pm0.0040.862±0.0040.862 \pm0.0040.875±0.0040.875 \pm0.0040.935±0.0020.935 \pm0.002
Layers 3-4MLL23-adapted0.935±0.0090.935 \pm0.0090.889±0.0100.889 \pm0.0100.872±0.0080.872 \pm0.0080.879±0.0080.879 \pm0.0080.936±0.0080.936 \pm0.008
Layer 4 onlyImageNet0.921±0.0030.921 \pm0.0030.866±0.0200.866 \pm0.0200.835±0.0090.835 \pm0.0090.847±0.0090.847 \pm0.0090.925±0.0030.925 \pm0.003
Layer 4 onlyMLL23-adapted0.929±0.0030.929 \pm0.0030.869±0.0070.869 \pm0.0070.849±0.0030.849 \pm0.0030.858±0.0020.858 \pm0.0020.930±0.0030.930 \pm0.003
FrozenImageNet0.786±0.0020.786 \pm0.0020.717±0.0070.717 \pm0.0070.681±0.0040.681 \pm0.0040.685±0.0030.685 \pm0.0030.802±0.0020.802 \pm0.002
FrozenMLL23-adapted0.869±0.0100.869 \pm0.0100.831±0.0250.831 \pm0.0250.776±0.0110.776 \pm0.0110.786±0.0160.786 \pm0.0160.881±0.0090.881 \pm0.009

Table S1. Transfer-learning comparison using ResNet34 models with affine classifier heads. Values are the mean ±\pm sample standard deviation of pooled OOF metrics across three seeds.

The frozen-backbone confusion matrices show higher mean recall for all seven classes, with the largest improvements for the two least-supported classes (Figure 3b). Mean recall increases from 46.3% to 74.1% for basophils (BAS, n=18n = 18) and from 50.0% to 68.9% for atypical lymphocytes (LYA, n=30n = 30). Mean precision and F1 also increase for all seven classes (Table S2).

ClassSupportInitialisationPrecisionRecallF1
BAS18ImageNet0.229±0.0110.229 \pm0.0110.463±0.0320.463 \pm0.0320.307±0.0150.307 \pm0.015
BAS18MLL23-adapted0.408±0.0300.408 \pm0.0300.741±0.0850.741 \pm0.0850.526±0.0400.526 \pm0.040
EOS43ImageNet0.819±0.0130.819 \pm0.0130.806±0.0130.806 \pm0.0130.812±0.0120.812 \pm0.012
EOS43MLL23-adapted0.939±0.0120.939 \pm0.0120.829±0.0270.829 \pm0.0270.881±0.0100.881 \pm0.010
KSC158ImageNet0.941±0.0340.941 \pm0.0340.916±0.0130.916 \pm0.0130.928±0.0100.928 \pm0.010
KSC158MLL23-adapted0.976±0.0040.976 \pm0.0040.960±0.0130.960 \pm0.0130.968±0.0090.968 \pm0.009
LYA30ImageNet0.230±0.0150.230 \pm0.0150.500±0.0330.500 \pm0.0330.314±0.0070.314 \pm0.007
LYA30MLL23-adapted0.329±0.0400.329 \pm0.0400.689±0.0510.689 \pm0.0510.445±0.0470.445 \pm0.047
LYT224ImageNet0.819±0.0220.819 \pm0.0220.731±0.0370.731 \pm0.0370.772±0.0180.772 \pm0.018
LYT224MLL23-adapted0.914±0.0300.914 \pm0.0300.815±0.0330.815 \pm0.0330.862±0.0220.862 \pm0.022
MON180ImageNet0.825±0.0180.825 \pm0.0180.781±0.0220.781 \pm0.0220.802±0.0110.802 \pm0.011
MON180MLL23-adapted0.922±0.0170.922 \pm0.0170.891±0.0250.891 \pm0.0250.906±0.0050.906 \pm0.005
NGS168ImageNet0.908±0.0180.908 \pm0.0180.821±0.0210.821 \pm0.0210.862±0.0130.862 \pm0.013
NGS168MLL23-adapted0.945±0.0040.945 \pm0.0040.889±0.0120.889 \pm0.0120.916±0.0080.916 \pm0.008

Table S2. Class-specific performance with and without intermediate MLL23 adaptation for ResNet34 models with a frozen backbone and an affine classifier head. Values are the mean ±\pm sample standard deviation across three seeds. For each seed, five-fold out-of-fold (OOF) predictions are pooled for all 821 images. Metrics are calculated from the raw counts in the resulting confusion matrix. Support denotes the number of images per class.

With MLL23-adapted initialisation, fine-tuning residual layers 3-4 yields a mean macro-F1 of 0.879, compared with 0.881 for full fine-tuning. Given this small difference, we use the layers 3-4 configuration for the head comparison, allowing the later features to adapt to each classifier while keeping the stem and residual layers 1-2 frozen.

Affine-hybrid comparison

Among the seven hybrid configurations evaluated using a single seed, the shallow two-block head B2/L1B2/L1 achieves the highest macro-F1 and balanced accuracy (Table 3). We therefore include it in the comparison with the affine head, alongside B1/L1B1/L1 as a single-block reference. Both hybrid heads are also evaluated using two additional seeds.

EncodingConfig.Aggregate metricsBASLYA
Macro-F1Balanced accuracyPrecisionRecallPrecisionRecall
Single encodingB1/L1B1/L10.8660.8730.8240.7780.4150.567
Single encodingB2/L1B2/L10.8690.8880.7500.8330.4350.667
Single encodingB3/L1B3/L10.8540.8650.7500.8330.3850.500
Standard re-uploadingB1/L2B1/L20.8040.8580.4520.7780.2920.700
Incremental re-uploadingB1/L2B1/L20.8300.8470.6670.6670.3270.600
Incremental re-uploadingB1/L3B1/L30.8000.8330.5000.7220.2580.533
Incremental re-uploadingB2/L2B2/L20.8400.8700.6150.8890.3400.533

Table 3. Macro-F1, balanced accuracy, and class-specific precision and recall for seven MLL23-adapted hybrid configurations evaluated using a single seed. All values are pooled OOF estimates for the same 821 images. Configuration labels specify the number of independent 12-qubit blocks (BB) and variational layers per block (LL). Single encoding denotes one input-encoding step per block (L=1L = 1).

The affine head achieves the highest mean macro-F1 (0.879 ± 0.008), followed by B2/L1B2/L1 (0.865 ± 0.010) and B1/L1B1/L1 (0.854 ± 0.013; Figure 4a). Balanced accuracy is numerically similar for the affine head and B2/L1B2/L1, at 0.889 ± 0.010 and 0.886 ± 0.008, respectively (Figure 4b). Table S3 reports the full aggregate results, including accuracy, macro precision, and weighted F1.

Comparison of the affine head and the two selected hybrid heads

Figure 4. Comparison of the affine head and the two selected hybrid heads. (a) Pooled OOF macro-F1. (b) Balanced accuracy. (c) Class-specific precision, recall, and F1 for BAS (n=18n = 18), LYA (n=30n = 30), and EOS (n=43n = 43). Bar labels show means across seeds 9, 122, and 250, and error bars show standard deviations. Metrics are calculated from pooled OOF predictions separately for each seed. All models start from the same MLL23-adapted ResNet34 weights and fine-tune residual layers 3-4 jointly with the classifier head

Table S3. Aggregate affine-hybrid comparison using the MLL23-adapted ResNet34 backbone with residual layers 3-4 fine-tuned. In the configuration labels, BB denotes the number of independent 12-qubit blocks and LL the number of variational layers per block. Values are the mean ±\pm sample standard deviation of pooled OOF metrics across three seeds.

HeadAccuracyBalanced AccuracyMacro-PrecisionMacro-F1Weighted-F1
Affine0.935±0.0090.935 \pm0.0090.889±0.0100.889 \pm0.0100.872±0.0080.872 \pm0.0080.879±0.0080.879 \pm0.0080.936±0.0080.936 \pm0.008
Hybrid B1/L1B1/L10.918±0.0110.918 \pm0.0110.876±0.0080.876 \pm0.0080.842±0.0200.842 \pm0.0200.854±0.0130.854 \pm0.0130.922±0.0100.922 \pm0.010
Hybrid B2/L1B2/L10.924±0.0090.924 \pm0.0090.886±0.0080.886 \pm0.0080.850±0.0100.850 \pm0.0100.865±0.0100.865 \pm0.0100.928±0.0080.928 \pm0.008

The low-support classes show different precision-recall trade-offs (Figure 4c and Table S4). Compared with the affine head, B2/L1B2/L1 has slightly higher mean recall for LYA (0.589 versus 0.567) and EOS (0.984 versus 0.977), while BAS recall remains at 0.870. For LYA, higher recall comes with lower precision (0.423 versus 0.478) and F1 (0.490 versus 0.517). BAS F1 is also lower (0.784 versus 0.832), whereas EOS F1 remains approximately 0.977. B1/L1B1/L1 has the same mean LYA recall as the affine head but lower recall for BAS and EOS. The recall gains observed for B2/L1B2/L1 therefore do not translate into a consistent F1 improvement across these classes.

Table S4. Pooled OOF precision, recall, and F1 for the three least-supported classes: basophils (BAS, n=18n = 18), atypical lymphocytes (LYA, n=30n = 30), and eosinophils (EOS, n=43n = 43). All pipelines use the MLL23-adapted ResNet34 backbone with residual layers 3-4 fine-tuned. Both hybrid heads use L=1L = 1, with independently simulated 12-qubit blocks. Values are the mean ±\pm sample standard deviation across three seeds.

ClassHeadPrecisionRecallF1
BASAffine0.798±0.0410.798 \pm0.0410.870±0.0320.870 \pm0.0320.832±0.0290.832 \pm0.029
BASHybrid B1/L1B1/L10.710±0.1030.710 \pm0.1030.852±0.0850.852 \pm0.0850.768±0.0470.768 \pm0.047
BASHybrid B2/L1B2/L10.715±0.0430.715 \pm0.0430.870±0.0320.870 \pm0.0320.784±0.0200.784 \pm0.020
LYAAffine0.478±0.0760.478 \pm0.0760.567±0.0580.567 \pm0.0580.517±0.0630.517 \pm0.063
LYAHybrid B1/L1B1/L10.385±0.0700.385 \pm0.0700.567±0.1000.567 \pm0.1000.458±0.0810.458 \pm0.081
LYAHybrid B2/L1B2/L10.423±0.0700.423 \pm0.0700.589±0.0690.589 \pm0.0690.490±0.0600.490 \pm0.060
EOSAffine0.977±0.0000.977 \pm0.0000.977±0.0000.977 \pm0.0000.977±0.0000.977 \pm0.000
EOSHybrid B1/L1B1/L10.970±0.0260.970 \pm0.0260.969±0.0130.969 \pm0.0130.969±0.0060.969 \pm0.006
EOSHybrid B2/L1B2/L10.970±0.0130.970 \pm0.0130.984±0.0130.984 \pm0.0130.977±0.0110.977 \pm0.011

Seed-specific confusion matrices are provided in Figure S1-Figure S3. For the seed used in the architectural exploration, B2/L1B2/L1 correctly classifies 20 of the 30 LYA images, compared with 16 for the affine head, but produces 26 rather than 15 false-positive LYA predictions (Figure S3).

Pooled OOF confusion matrices for seed 9, showing Affine, Hybrid B1/L1, and Hybrid B2/L1 classifier heads

Figure S1. Pooled OOF confusion matrices for seed 9. From left to right: Affine, Hybrid B1/L1B1/L1, and Hybrid B2/L1B2/L1. Cell annotations indicate image counts, whereas colour represents percentages within each true class.

Pooled OOF confusion matrices for seed 250

Figure S3. Pooled OOF confusion matrices for seed 250. From left to right: Affine, Hybrid B1/L1B1/L1, and Hybrid B2/L1B2/L1. Cell annotations indicate image counts, whereas colour represents percentages within each true class

Encoding strategy, variational layer count, and independent blocks

Table 3 compares seven hybrid configurations evaluated using a single seed. With single encoding, B2/L1B2/L1 achieves the highest macro-F1 (0.869) and balanced accuracy (0.888), followed by B1/L1B1/L1 and B3/L1B3/L1. Adding blocks also increases the size of the classical projection and readout.

For B1/L2B1/L2, incremental re-uploading yields higher macro-F1 than standard re-uploading (0.830 versus 0.804), but lower balanced accuracy (0.847 versus 0.858). BAS and LYA both show higher precision and lower recall.

For the single-block heads, macro-F1 decreases from 0.866 with single encoding at L=1L = 1 to 0.804 with standard re-uploading at L=2L = 2. With incremental re-uploading, increasing LL from two to three reduces macro-F1 from 0.830 to 0.800. Balanced accuracy also decreases in both comparisons.

Under noiseless statevector simulation, B1/L1B1/L1 and B2/L1B2/L1 require approximately 101 and 198 times the training time of the affine baseline. Table S5 reports runtimes, trainable head parameter counts, and measurement conditions.

Discussion

Dataset adaptation and feature learning The MLL23 adaptation improves the mean macro-F1 score across all four fine-tuning configurations, with the greatest improvement observed when the backbone is totally frozen. Since only the affine head learns from the target images in this setting, this suggests that the adapted features better distinguish the target classes. Recall improves across all seven classes, particularly BAS and LYA. On the other hand, the smaller gains after fine-tuning may indicate that target-task learning can partly compensate for differences in the initial features.

The comparison between the different heads is based on the hybrid transfer-learning framework described in Mari et al. [16], in which the subsequent layers of the backbone are trained jointly with each head. The sharing of initial weights does not imply that the final features are identical: the results also reflect the way in which the base network adapts to each classifier.

Beyond aggregate performance. The similar balanced accuracy of the affine and two-block heads masks differences in their errors. For B2/L1, which uses two independent 12-qubit blocks with one variational layer each, higher mean LYA recall comes with lower precision and F1. Class-specific analysis therefore helps to determine whether an improvement in recall leads to an increase in false positives.

The high F1 score for EOS, despite its limited support, may partly reflect the characteristic cytoplasmic granules of eosinophils [26]. Atypical lymphocytes, by contrast, have a variable appearance and may resemble typical lymphocytes or monocytes [18]. This morphological overlap would explain the frequent confusion of LYA with LYT and MON, but makes the classification problem particularly challenging when there are few training samples.

Circuit depth. Repeated encoding can enlarge the frequency spectrum accessible to quantum models [25], but greater representational capacity does not guarantee better classification. The deeper single-block configurations have lower aggregate scores and longer training times.

Encoding and classical components. The performance of B2/L1 in the architectural exploration motivates further study of independent blocks, while the encoding comparison shows that gains depend on the metric considered. These architectural choices also change the classical components: incremental re-uploading enlarges the projection, and additional blocks enlarge both the projection and readout. Cruz et al. [6] kept the surrounding architecture fixed when comparing intermediate transformations and reported a higher macro-F1 score for hybrid models. Our study evaluates complete head replacement following MLL23 adaptation, with a particular focus on a small, imbalanced reference dataset that distinguishes between typical and atypical lymphocytes.

Limitations. One of the limitations of the evaluation protocol concerns the overlap between model development and performance evaluation. This choice was motivated by the limited number of available samples, particularly in the least-supported classes. Performance on the validation folds was used to adjust learning rates and select the epoch with the best performance, whilst the development of the architecture was based on the same benchmark dataset used for evaluation. The aggregate OOF estimates may therefore be influenced by selection bias [32, 4].

Conclusion

This study shows how domain adaptation and classifier-head design affect blood-cell classification on an imbalanced microscopy dataset. Intermediate MLL23 adaptation improves mean macro-F1 across the four affine-head fine-tuning settings, with the largest gain when the backbone is frozen. With the MLL23-adapted backbone, the shallow two-block hybrid head achieves mean macro-F1 and balanced accuracy scores that are numerically comparable to those of the affine head, given the associated uncertainties. The comparison also reveals distinct precision-recall trade-offs across classes.

The architectural exploration favours shallow configurations and shows that encoding and block choices are associated with different patterns of classification errors. Taken together, these results provide practical guidance for selecting hybrid classifiers based on both overall performance and the trade-off between precision and recall for under-represented classes. Future work could extend the hybrid approach by integrating quantum components into feature extraction and by exploring trainable quantum pre-processing modules. Joint optimisation of these components with the classifier could clarify whether quantum pre-processing improves the ability to distinguish between morphologically similar cell types, particularly those for which training data is limited.

Statements and Declarations

Funding This work was carried out as part of the SecQDevOps project (Trustworthy Quantum DevOps for Secure Software and Hardware Quantum Engineering), funded by the European Union under the Horizon Europe research and innovation programme (Grant Agreement No. 101225776). This work has received funding from the Swiss State Secretariat for Education, Research and Innovation (SERI) through its support for HES-SO’s participation in SecQDevOps (project No. 25.00261).

Competing interests The authors declare no competing interests.

Author contributions G.D., E.P., J.E., E.S., and G.N. conceived the study. G.D. and C.A. developed the methodology and software and conducted the experiments. G.D. curated and analysed the data, prepared the figures, and drafted the manuscript. G.D., G.N., E.S., S.F.C., J.E., and E.P. reviewed and edited the manuscript. J.E. and E.P. supervised the project. All authors approved the manuscript for submission.

Ethics approval This study uses existing images from the Raw-Microscopy and MLL23 datasets. Raw-Microscopy comprises images acquired from commercially obtained, de-identified human blood smear slides [19]. The original MLL23 work received approval from the Ethics Committee of LMU Munich (reference number 25-0744) [29]. No new biological samples were collected for this study.

Data and code availability The Raw-Microscopy dataset and associated data models are described by Oala et al. [19] and archived on Zenodo [20]. The MLL23 dataset is described by Shetab Boushehri et al. [29] and available on Zenodo [28].

The implementation is available on GitHub at https://github.com/gabrieledangeli/hybrid-quantum-classical-microscopy-classification. The associated Weights & Biases project is available at https://wandb.ai/gabriele-dangeli2-hes-so-secqdevops/MicroscopyRGB_Classification. The filenames of the 81 duplicate images removed from MicroscopyRGB and the MLL23-adapted ResNet34 backbone weights used in this study, together with checkpoint metadata and loading instructions, are publicly available on Zenodo [8].

Acknowledgements

We thank Nicolas Brunner and Pavel Sekatski for stimulating discussions about the quantum information aspects of this work, and Michèle Buck and Christian Matek for helpful discussions on data labelling.

We used ChatGPT (OpenAI) to improve the grammar and style of the manuscript. The authors reviewed and edited all suggested revisions and take full responsibility for the final content.

Aggregate and class-specific performance

Tables S1-S4 report aggregate and class-specific results for the transfer-learning and classifier-head comparisons. For each seed (9, 122, and 250), the evaluation metrics are calculated from pooled five-fold OOF predictions obtained at the epoch with the highest validation macro-F1 in each fold. Each of the 821 images contributes one prediction.

Table S1: Transfer-learning comparison using ResNet34 models with affine classifier heads. Values are the mean ±\pm sample standard deviation of pooled OOF metrics across three seeds.

Table S2: Class-specific performance with and without intermediate MLL23 adaptation for ResNet34 models with a frozen backbone and an affine classifier head.

Detailed training costs

Table S5. Observed end-to-end training runtimes and trainable head parameter counts for the affine baseline and seven hybrid configurations. Each runtime is the total time recorded in Weights & Biases across the five cross-validation folds for seed 250. All configurations are trained on a single NVIDIA A40 GPU for 50 epochs per fold. Hybrid circuits are evaluated by noiseless statevector simulation.

ConfigurationTrainable head params.CV runtime (h)
Affine3 5910.224
Single encoding B1/L1B1/L119 94322.716
Single encoding B2/L1B2/L138 85544.333
Single encoding B3/L1B3/L157 76766.698
Standard re-uploading B1/L2B1/L220 02734.613
Incremental re-uploading B1/L2B1/L238 49534.519
Incremental re-uploading B1/L3B1/L357 04747.172
Incremental re-uploading B2/L2B2/L275 95970.354

Seed-specific confusion matrices

Figures S1–S3 present the pooled OOF confusion matrices for each seed and classifier head. All three pipelines use the MLL23-adapted ResNet34 backbone with residual layers 3–4 fine-tuned.

Rows represent true classes and columns represent predicted classes.

Pooled OOF confusion matrices for seed 122, showing Affine, Hybrid B1/L1, and Hybrid B2/L1 classifier heads

Figure S2. Pooled OOF confusion matrices for seed 122. From left to right: Affine, Hybrid B1/L1B1/L1, and Hybrid B2/L1B2/L1. Cell annotations indicate image counts, whereas colour represents percentages within each true class.

References

  1. [1]Biewald L (2020) Experiment tracking with Weights and Biases. URL https://www.wandb.com/, software available from wandb.com. Accessed 10 November 2026.
  2. [2]Bowles J, Ahmed S, Schuld M (2024) Better than classical? The subtle art of benchmarking quantum machine learning models. https://doi.org/10.48550/arXiv.2403.07059, arXiv:2403.07059
  3. [3]Caro MC, Huang HY, Cerezo M, et al (2022) Generalization in quantum machine learning from few training data. Nature Communications 13(1):4919. https://doi.org/10.1038/s41467-022-32550-3
  4. [4]Cawley GC, Talbot NLC (2010) On over-fitting in model selection and subsequent selection bias in performance evaluation. Journal of Machine Learning Research 11(70):2079–2107. URL https://jmlr.org/papers/v11/cawley10a.htmlDOI
  5. [5]Chan S, Kulkarni P, Yi PH, et al (2024) Expanding the horizon: Enabling hybrid quantum transfer learning for long-tailed chest x-ray classification. In: 2024 IEEE International Conference on Quantum Computing and Engineering (QCE). IEEE, pp 572–582, https://doi.org/10.1109/QCE60285.2024.00073
  6. [6]Cruz G, Innan N, Marchisio A, et al (2026) Enhancing blood cells classification using hybrid quantum neural networks. https://doi.org/10.48550/arXiv.2605.23324, accepted at the IEEE International Conference on Quantum Computing and Engineering (QCE), 2026, arXiv:2605.23324
  7. [7]Cui Y, Jia M, Lin TY, et al (2019) Class-balanced loss based on effective number of samples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp 9268–9277, https://doi.org/10.1109/CVPR.2019.00949
  8. [8]D’Angeli G (2026) MicroscopyRGB duplicate list and MLL23-adapted ResNet34 backbone. Zenodo, https://doi.org/10.5281/zenodo.24141386, version 1.0.0
  9. [9]Deng J, Dong W, Socher R, et al (2009) ImageNet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp 248–255, https://doi.org/10.1109/CVPR.2009.5206848
  10. [10]Gupta RS, Wood CE, Engstrom T, et al (2025) A systematic review of quantum machine learning for digital health. npj Digital Medicine 8(1):237. https://doi.org/10.1038/s41746-025-01597-z
  11. [11]He K, Zhang X, Ren S, et al (2016) Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp 770–778, https://doi.org/10.1109/CVPR.2016.90
  12. [12]Kandala A, Mezzacapo A, Temme K, et al (2017) Hardware-efficient variational quantum eigensolver for small molecules and quantum magnets. Nature 549(7671):242–246. https://doi.org/10.1038/nature23879
  13. [13]Koch V, Wagner SJ, Kazeminia S, et al (2024) DinoBloom: A foundation model for generalizable cell embeddings in hematology. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, Lecture Notes in Computer Science, vol 15012. Springer Nature Switzerland, pp 520–530, https://doi.org/10.1007/978-3-031-72390-2_49
  14. [14]Landman J, Mathur N, Li YY, et al (2022) Quantum methods for neural networks and application to medical image classification. Quantum 6:881. https://doi.org/10.22331/q-2022-12-22-881
  15. [15]Loshchilov I, Hutter F (2019) Decoupled weight decay regularization. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=Bkg6RiCqY7
  16. [16]Mari A, Bromley TR, Izaac J, et al (2020) Transfer learning in hybrid classical-quantum neural networks. Quantum 4:340. https://doi.org/10.22331/q-2020-10-09-340
  17. [17]Matic A, Monnet M, Lorenz JM, et al (2022) Quantum-classical convolutional neural networks in radiological image classification. In: 2022 IEEE International Conference on Quantum Computing and Engineering (QCE). IEEE, pp 56–66, https://doi.org/10.1109/QCE53715.2022.00024
  18. [18]Nozaka H, Kushibiki M, Kamata K, et al (2023) Classifying microscopic images of reactive lymphocytosis using two-step tandem AI models. Applied Sciences 13(9):5296. https://doi.org/10.3390/app13095296
  19. [19]Oala L, Aversa M, Nobis G, et al (2023a) Data models for dataset drift controls in machine learning with optical images. Transactions on Machine Learning Research URL https://openreview.net/forum?id=I4IkGmFgJzarxiv.org/abs/2211.02578
  20. [20]Oala L, Aversa M, Nobis G, et al (2023b) Data models for dataset drift controls in machine learning with optical images - datasets. Zenodo, https://doi.org/10.5281/zenodo.5235536, dataset, version 1.0.0-beta
  21. [21]Pedregosa F, Varoquaux G, Gramfort A, et al (2011) Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12(85):2825–2830. URL https://jmlr.org/papers/v12/pedregosa11a.htmlarxiv.org/abs/1201.0490
  22. [22]Pérez-Salinas A, Cervera-Lierta A, Gil-Fuster E, et al (2020) Data re-uploading for a universal quantum classifier. Quantum 4:226. https://doi.org/10.22331/q-2020-02-06-226
  23. [23]Periyasamy M, Meyer N, Ufrecht C, et al (2022) Incremental data-uploading for full-quantum classification. In: 2022 IEEE International Conference on Quantum Computing and Engineering (QCE). IEEE, pp 31–37, https://doi.org/10.1109/QCE53715.2022.00021
  24. [24]Russakovsky O, Deng J, Su H, et al (2015) ImageNet large scale visual recognition challenge. International Journal of Computer Vision 115(3):211–252. https://doi.org/10.1007/s11263-015-0816-y
  25. [25]Schuld M, Sweke R, Meyer JJ (2021) Effect of data encoding on the expressive power of variational quantum-machine-learning models. Physical Review A 103(3):032430. https://doi.org/10.1103/PhysRevA.103.032430
  26. [26]Scordino T (2016) Eosinophil. American Society of Hematology Image Bank, URL https://imagebank.hematology.org/image/60934/eosinophil?type=atlas, image 60934. Accessed 10 November 2026.
  27. [27]Senokosov A, Sedykh A, Sagingalieva A, et al (2024) Quantum machine learning for image classification. Machine Learning: Science and Technology 5(1):015040. https://doi.org/10.1088/2632-2153/ad2aef
  28. [28]Shetab Boushehri S, Gruber A, Sun X, et al (2023) A large publicly available single-cell peripheral blood dataset (MLL23). Zenodo, https://doi.org/10.5281/zenodo.14277609, dataset
  29. [29]Shetab Boushehri S, Kazeminia S, Gruber A, et al (2025) A large expert-annotated single-cell peripheral blood dataset for hematological disease diagnostics. Scientific Data 12(1):1773. https://doi.org/10.1038/s41597-025-06223-x
  30. [30]Sünkel L, Altmann P, Kölle M, et al (2024) On the quantum impact in hybrid classical-quantum transfer learning. In: 2024 IEEE International Conference on Quantum Computing and Engineering (QCE), pp 11–15, https://doi.org/10.1109/QCE60285.2024.10244
  31. [31]Umer RM, Gruber A, Shetab Boushehri S, et al (2023) Imbalanced domain generalization for robust single cell classification in hematological cytomorphology. In: ICLR 2023 Workshop on Domain Generalization, URL https://openreview.net/pdf?id=JWD0ioscQEarxiv.org/abs/2303.07771
  32. [32]Varma S, Simon R (2006) Bias in error estimation when using cross-validation for model selection. BMC Bioinformatics 7:91. https://doi.org/10.1186/1471-2105-7-91

Paper details

vlnc:W_btc72pdj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

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

  • Engineering

ChatGPT (OpenAI)

View statement
Looking for related papers…
Contents