betaFeedback

WorksPhysics and Astronomy

Distinguishing thyroid nodule malignancy using digital palpation

Abstract

Accurate pre-operative diagnosis of thyroid nodules remains a major clinical challenge. Current diagnostic techniques rely on fine-needle aspiration and demonstrate high rates of inconclusive results, often leading to avoidable thyroidectomies with lifelong implications. Here, we introduce a mechanics-based diagnostic method that leverages Digital Palpation to determine malignancy and identify tissue subtypes in human thyroid nodules. From a cohort of patients undergoing thyroidectomies, we evaluated ex vivo thyroid specimens and processed the resulting mechanical data from 36 tests in 10 thyroid nodules. Principal Component Analysis reveals that these experimental results are distributed within a three-dimensional space of interpretable scalar features reflecting tissue grittiness, strength, and toughness, which map closely to descriptions of malignant and benign nodules given by clinicians. Our results further demonstrate that malignant nodules exhibit distinct mechanical signatures compared to their benign counterparts. This revealed a statistically significant difference (p = 9 x 10^-4) and separability (95% lower-bound Hedges' g = 1.08) between benign and malignant nodules along a direction in principal component space representing effective stiffness. Preliminary clustering patterns were also observed among malignant subtypes, including follicular, oncocytic, and papillary thyroid carcinomas. Together, these results highlight the potential of Digital Palpation as a novel mechanics-based method for tissue characterization and motivate its further development toward diagnostic applications.

Introduction

By the age of 60, the majority of adults will present with a thyroid nodule [1, 2, 3], i.e. a discrete lesion within the thyroid gland that is radiologically distinct from surrounding thyroid tissues [4]. While most nodules (∼90%\sim90\%) are benign [5], some may harbor cancer. In 2022, there were an estimated 1,016,930 people living with thyroid cancer (TC) in the United States [6], reflecting a significant increase in the incidence of TC, which has tripled in the past three decades [7, 8, 9, 10, 11]. In the same year, TC was reported as the fifth most commonly diagnosed cancer worldwide in women; particularly affecting young women (<50< 50 years), for which it was the second most commonly diagnosed cancer [7, 12].

The increasing prevalence of TC explains an increase in the need for reliable and minimally invasive diagnosis methods for TC. The existing standard of care for TC diagnostics relies on fine-needle aspiration (FNA) followed by cytology, which has a broad range of reported false negative rates (3-46%) that increases with nodule size [13, 14, 15, 16, 17, 18], and cannot decisively classify up to one third of nodules [19, 20, 21, 22, 23], leading patients to undergo molecular testing, and then diagnostic surgery, i.e. a thyroidectomy. As a result, a large fraction of resected thyroids (40-70%) ultimately receive a benign pathology result [24, 25, 26], implying that the major surgical intervention may have been avoidable [27, 28, 29]. Regardless of the final pathology result, the increasing population of patients being treated for TC results in substantial costs to patients, insurers, and society [30, 31]. Patients that are treated for TC experience significant effects to their quality of life on par with other cancers [30, 32, 33]. Including the need for life-long hormone replacement and the management of potential surgical complications that may include vocal cord paralysis and hypocalcemia [34, 35, 36, 37, 38].

These shortcomings have motivated efforts to improve upon the existing diagnostic methods and the current standard of care. One significant improvement has been achieved by the development of molecular testing (e.g. Afirma and Throseq) [39, 40, 41, 42, 43], which was shown to reduce the number of diagnostic surgeries in indeterminate nodules [44]; however, molecular testing relies on successful FNA and remains an expensive and time intensive (2-3 weeks) diagnostic. Notably, some studies report that, in practice, many patients who receive molecular testing eventually undergo surgical resection [45, 46, 47, 48, 49, 50]. Another approach to improve TC diagnosis has leveraged advanced imaging based methods [51, 52, 53] (e.g. elastography, PETCT); however, these approaches have not shown improved accuracy over the current standard of care [9, 53, 54, 55, 56, 57], and are thus not currently recommended in the diagnosis workflow [49]. Similarly, the use of Artificial Intelligence (AI) to enhance imaging based disease classification has yet to offer improved accuracy over the current standard of care [58, 59].

Motivated by the common perception among clinicians that various malignant and benign nodules are occasionally physically distinguishable to the touch, a nascent and promising approach aims to leverage these reported differences to improve the classification of nodules. After handling the tissue during surgery, in the pathology laboratory, or while performing FNA, clinicians often use terms such as tough, gritty, firm, hard, etc. [60, 61, 62], to describe and distinguish nodules. Notably, clinicians often specify that malignant nodules feel particularly stiff. Studies have shown that such mechanical indications obtained via manual palpation [63] add clinical value and can serve as auxiliary diagnostic evidence [62, 64, 65, 66]; however, the highly subjective and qualitative nature of these clinical descriptions limit their diagnostic value. Accordingly, if we could quantify these mechanical descriptions of nodules, diagnosis could improve.

To quantify mechanical heterogeneity in human thyroid tissue lesions Sharma et al. [66] employed piezoelectric needle sensors that measure the resisting force upon needle insertion. They reported distinguishable force and heterogeneity values for different tissue types, providing a first indication of mechanically measurable differences; however, the use of force measurements directly behind the needle creates a dependence on the geometry of the nodule and proximity to other tissues (which as such must be known prior to ascertaining a definitive diagnosis in real clinical settings). Ultimately, to be able to provide robust and repeatable results, a mechanical test for distinguishing and diagnosing thyroid cancer must provide a local measure of material response within the nodule that is independent of the macroscopic geometry of the thyroid.

In this study, we show that Digital Palpation; a method to quantitatively characterize the local mechanical resistance of a nodule, independent of its macroscopic geometry or physical location; can distinguish freshly excised human thyroid nodules. This offers a new pathway for improved diagnosis and risk stratification of thyroid nodule malignancies. By measuring the resisting pressure developed within the water column in a sharp injection needle, as it is driven through the tissue layers and ultimately into the nodule, the tearing resistance (amongst other properties) of the tissue is captured. Then by expanding a nanoliter-sized fluid bubble with concurrent pressure measurement, the mechanical resistance to cavity expansion is measured. This multivariate mechanical profile from 6 thyroid specimens, obtained from 36 tests in 10 nodules, was examined via principal component analysis (PCA) [67] and logistic regression. Signed Euclidean distances from the logistic regression decision boundary were then computed to visualize dimensionality reduction and highlight clusters between benign and malignant nodules, as well as among malignancy subtypes including follicular (FTC), oncocytic (OTC), and papillary thyroid carcinoma (PTC). We also note that the signed Euclidian distance measure separating malignant and benign nodules aligns with clinical descriptions of stiffer malignant nodules from manual palpation without assuming so a priori.

Results and Discussion

Digital Palpation (DP) enables collection of mechanical data in freshly excised human thyroids

The reports from physicians mentioned above include various terminologies to describe complex differences between nodule types obtained via manual palpation or during FNA testing. Although these reports motivate our effort to quantitatively distinguish nodules via their mechanical response, these reported differences cannot be accurately described using conventional mechanical measures, such as stiffness or toughness alone. Hence, in this work, we depart from common experimental approaches to study the mechanical response of biological tissue, which is usually geared towards the development of constitutive models. Instead, we aim to utilize the mechanical data to maximally distinguish and classify tissue types in a rapid and minimally invasive way. To this end, we developed a method to quantitatively interrogate the local mechanical resistance of the tissue, at the tip of a needle, throughout both needle insertion and cavity expansion.

Our approach, which we refer to as Digital Palpation (DP), builds on and combines two recently developed methods: needle-based insertion [68], and volume-controlled cavity expansion [69, 70] (VCCE); both of which are designed to obtain objective mechanical data that is independent of the macroscopic geometry. Notably, needle-based insertion has been shown to reliably distinguish between soft synthetic materials [68] and to inform best needle insertion protocols, and VCCE has been successfully applied to characterize not only synthetic materials, but also a variety of biological materials; showing a robust ability to distinguish gray matter from white matter brain tissue [71], changes in blood clots coagulated on different days [72], and aging of liver [73], highlighting their potential for use in clinical settings.

As explained in greater detail in the Materials and Methods section, surgically excised thyroid specimens exhibiting thyroid nodules were evaluated at Mass General Brigham Newton-Wellesley Hospital using DP with IRB approval from the Mass General Brigham IRB. As seen in Figure 1a, freshly resected thyroid was placed on the DP translation stage, and positioned directly underneath the needle-plunger system. Controlled needle insertion was achieved by raising the translation stage (Figure 1b) at a constant rate of 12.50mm/s, targeting the center of the thyroid nodule of interest. The target insertion depth was informed by the overseeing clinician, taking into account the nodule’s distance from the thyroid surface and whether the specimen was repositioned or reoriented to avoid repeated insertions at similar locations (see Table S1 and Table S2 in the Supplementary). During insertion, the variation of pressure within the fluid column was recorded to produce the insertion curve as seen in the inset of Figure 1(b). Once insertion was complete, the injection of the phosphate buffered saline (PBS) water solution began at a volumetric flow rate of 0.6 μL/s0.6\,\mu\mathrm{L}/\mathrm{s} until a total fluid volume of 9.2 μL9.2\,\mu\mathrm{L} was transferred. This fluid volume is equivalent to a spherical cavity with final radius of 1.30 mm1.30\,\mathrm{mm}.

Schematic overview of the Digital Palpation testing setup and data collection protocol

Figure 1. Schematic overview of the Digital Palpation testing setup and data collection protocol. (a) Preparation: A fixed syringe assembly (I) is pre-filled with Phosphate-buffered saline and fitted with a pressures sensor (II), and a 25 Gauge hypodermic needle (III). A resected thyroid specimen containing nodules (inset) is positioned atop a vertically actuated translation stage (IV) with the interrogated nodule placed directly below the needle. (b) Insertion: Controlled needle insertion is achieved by driving the actuated stage towards the needle. Simultaneous recording of the pressure produces a pressure-time response referred to as the Needle Insertion Curve, as shown schematically (inset). We denote the peak pressure during insertion by c1c_1. (c) Injection: The fluid injection procedure is subsequently initiated by displacement of the syringe plunger within the syringe assembly, injecting fluid at a controlled rate (with nanoliter precision) into the tissue to expand a millimeter scale cavity. The resistance of the tissue to expansion results in an increase in pressure that is recorded via the pressure sensor to obtain a pressure-volume response termed the Cavity Expansion Curve, as shown schematically (inset). We identify three key characteristic features; the peak pressure c2c_2, the corresponding volume c3c_3, and lastly, c4c_4 corresponds to the number of peaks.

#T–NPre-Operative UltrasoundPost-Operative PathologyAP × TRV × SG (cm)δ\delta (mm)
11–1Bethesda II: Benignthyroid follicular nodular disease1×0.9×0.91 \times0.9 \times0.9 to 4.1×3.9×34.1 \times3.9 \times310
21–1Bethesda II: Benignthyroid follicular nodular disease1×0.9×0.91 \times0.9 \times0.9 to 4.1×3.9×34.1 \times3.9 \times310
31–2Bethesda II: Benignthyroid follicular nodular disease1×0.9×0.91 \times0.9 \times0.9 to 4.1×3.9×34.1 \times3.9 \times310
41–2Bethesda II: Benignthyroid follicular nodular disease1×0.9×0.91 \times0.9 \times0.9 to 4.1×3.9×34.1 \times3.9 \times315
51–3Bethesda II: Benignthyroid follicular nodular disease1.7×1.1×0.91.7 \times1.1 \times0.95
61–3Bethesda II: Benignthyroid follicular nodular disease1.7×1.1×0.91.7 \times1.1 \times0.99
71–3Bethesda II: Benignthyroid follicular nodular disease1.7×1.1×0.91.7 \times1.1 \times0.98
82–1Bethesda III: Indeterminatemacrofollicular nodule3.2×3.2×2.73.2 \times3.2 \times2.715
92–1Bethesda III: Indeterminatemacrofollicular nodule3.2×3.2×2.73.2 \times3.2 \times2.720
102–2Biopsy not performeddiffuse follicular hyperplasia0.8×0.8×0.70.8 \times0.8 \times0.78
112–2Biopsy not performeddiffuse follicular hyperplasia0.8×0.8×0.70.8 \times0.8 \times0.78
122–2Biopsy not performeddiffuse follicular hyperplasia0.8×0.8×0.70.8 \times0.8 \times0.78
132–2Biopsy not performeddiffuse follicular hyperplasia0.8×0.8×0.70.8 \times0.8 \times0.78
142–2Biopsy not performeddiffuse follicular hyperplasia0.8×0.8×0.70.8 \times0.8 \times0.715

Table S1. Summary of pre-operative ultrasound and post-operative pathology of benign thyroid (T) nodules (N). Approximate nodule size is represented as Anteroposterior (AP) × Transverse (TRV) × Sagittal (SG) dimensions. Additionally, total programmed injection depth (δ\delta) is shown. δ\delta varies with the depth of the nodule within the thyroid specimen and with specimen repositioning and reorientation to avoid repeated measurements at the same site. Note, δ\delta is reported for completeness and results within the paper do not depend on δ\delta.

Table S2: Summary of pre-operative ultrasound results and post-operative pathology for malignant thyroid (T) nodules (N). Approximate nodule size is represented as Anteroposterior (AP) ×\times Transverse (TRV) ×\times Sagittal (SG) dimensions, and including TNM staging and AJCC classification. Additionally, total programmed injection depth (δ\delta) is shown. δ\delta varies with the depth of the nodule within the thyroid specimen and with specimen repositioning and reorientation to avoid repeated measurements at the same site. Note, δ\delta is reported for completeness and results within the paper do not depend on δ\delta.

#T–NPre-Operative UltrasoundPost-Operative PathologyTNM & AJCCAP ×\times TRV ×\times SG (cm)δ\delta (mm)
13–1Bethesda III: Indeterminatefollicular thyroid carcinomaT1bNxMx; I1.6 ×\times 1.4 ×\times 0.95
23–1Bethesda III: Indeterminatefollicular thyroid carcinomaT1bNxMx; I1.6 ×\times 1.4 ×\times 0.95
33–1Bethesda III: Indeterminatefollicular thyroid carcinomaT1bNxMx; I1.6 ×\times 1.4 ×\times 0.95
43–2Bethesda IV: Indeterminatefollicular thyroid carcinomaT1bNxMx; I1.6 ×\times 1.4 ×\times 0.94
53–2Bethesda IV: Indeterminatefollicular thyroid carcinomaT1bNxMx; I1.6 ×\times 1.4 ×\times 0.94
64–1Bethesda VI: Malignantpapillary thyroid carcinomaT2N1aMx; II2.2 ×\times 1.1 ×\times 18
74–1Bethesda VI: Malignantpapillary thyroid carcinomaT2N1aMx; II2.2 ×\times 1.1 ×\times 16
84–1Bethesda VI: Malignantpapillary thyroid carcinomaT2N1aMx; II2.2 ×\times 1.1 ×\times 17
94–1Bethesda VI: Malignantpapillary thyroid carcinomaT2N1aMx; II2.2 ×\times 1.1 ×\times 18
104–1Bethesda VI: Malignantpapillary thyroid carcinomaT2N1aMx; II2.2 ×\times 1.1 ×\times 17
114–1Bethesda VI: Malignantpapillary thyroid carcinomaT2N1aMx; II2.2 ×\times 1.1 ×\times 18
124–1Bethesda VI: Malignantpapillary thyroid carcinomaT2N1aMx; II2.2 ×\times 1.1 ×\times 15
135–1Bethesda VI: Malignantpapillary thyroid carcinomaT1bN1aMx; II1.1 ×\times 1.0 ×\times 0.76
145–1Bethesda VI: Malignantpapillary thyroid carcinomaT1bN1aMx; II1.1 ×\times 1.0 ×\times 0.75
155–1Bethesda VI: Malignantpapillary thyroid carcinomaT1bN1aMx; II1.1 ×\times 1.0 ×\times 0.75
165–1Bethesda VI: Malignantpapillary thyroid carcinomaT1bN1aMx; II1.1 ×\times 1.0 ×\times 0.720
176–1Bethesda IV: Indeterminateoncocytic thyroid carcinomaT2NxMx; II3.7 ×\times 3.2 ×\times 2.620
186–1Bethesda IV: Indeterminateoncocytic thyroid carcinomaT2NxMx; II3.7 ×\times 3.2 ×\times 2.615
196–1Bethesda IV: Indeterminateoncocytic thyroid carcinomaT2NxMx; II3.7 ×\times 3.2 ×\times 2.613
206–1Bethesda IV: Indeterminateoncocytic thyroid carcinomaT2NxMx; II3.7 ×\times 3.2 ×\times 2.614
216–1Bethesda IV: Indeterminateoncocytic thyroid carcinomaT2NxMx; II3.7 ×\times 3.2 ×\times 2.613
226–1Bethesda IV: Indeterminateoncocytic thyroid carcinomaT2NxMx; II3.7 ×\times 3.2 ×\times 2.614

Table S2. Summary of pre-operative ultrasound results and post-operative pathology for malignant thyroid (T) nodules (N). Approximate nodule size is represented as Anteroposterior (AP) ×\times Transverse (TRV) ×\times Sagittal (SG) dimensions, and including TNM staging and AJCC classification. Additionally, total programmed injection depth (δ\delta) is shown. δ\delta varies with the depth of the nodule within the thyroid specimen and with specimen repositioning and reorientation to avoid repeated measurements at the same site. Note, δ\delta is reported for completeness and results within the paper do not depend on δ\delta.

Characteristic Features Emerge in DP Data

The data collected via the DP protocol described above may be seen in the Supplementary. From these datasets, we observe typical features that emerge in both the needle insertion and cavity expansion curves, as illustrated schematically in Figure 1. For the present analysis, we selected four key features {c1,c2,c3,c4}\{c_1, c_2, c_3, c_4\}. From the needle insertion curve, we extracted the peak pressure c1c_1 achieved over the entire insertion process, which may be thought of as the maximal resistance of the nodule to needle penetration. From the cavity expansion curves we selected the remaining three features. In a typical cavity expansion process, the pressure increases monotonically as fluid volume increases up until a peak value. Similar to the needle insertion curve, we extract the peak pressure c2c_2 during fluid injection and the corresponding injected fluid volume at this point c3c_3. These features represent the transition between elastic expansion to localized tissue rupture. Lastly, the number of discrete peaks that in the injection data, NN, is also recorded, which may be thought of as a measure of tissue granularity (or heterogeneity), i.e. the number of rupture events as the fluid cavity expands. Due to NN’s discrete nature and the existence of true zeroes, this feature is represented by the regularized quantity c4=N+1+N2c_4 = N + \sqrt{1 + N^2}. In what follows, we will show that these four scalar features {c1,c2,c3,c4}\{c_1, c_2, c_3, c_4\} provide sufficient mechanical data to classify thyroid nodules. Further details on the extraction of these features from the needle insertion and cavity expansion curves may be found in the Materials and Methods section of the Supplementary. Importantly, no leaks or nodule targeting issues were identified in any of the tests. Hence, none of the tests met criteria for exclusion and all data recorded with the aforementioned needle insertion and cavity expansion protocol were used in the analysis.

In Figure 2(a,Figure 2b) we examine the differences between data that corresponds to nodules that later received a malignant pathology result (red), from data of those that were later found to be benign (blue). For both needle insertion and cavity expansion the corresponding averaged curves are shown via solid lines and the shaded regions represent inter-quartile range. While qualitative differences between groups are visually apparent, substantial overlap does exist, and the variability in each dataset is large. These differences thus motivate the use of statistical methods to better quantify and classify these results.

Raw DP data processed to obtain normalized features and principal components

Figure 2. (a) Needle insertion data, and (b) cavity expansion data, differentiated between benign (blue) and malignant (red) nodules. In both graphs, the solid lines correspond to the median-averaged value of each group and the shaded regions correspond to the inter-quartile range, respectively, to give an idea of variability between tests. (c) Extracted features cic_i are first log-transformed (Ci=ln⁡(ci)C_i = \ln(c_i)) and then z-score normalized, such that each feature distribution for benign (red) and malignant (blue) nodules has zero mean and unit standard deviation. (d-f) Principal component analysis results. Feature loadings for the first three principal components illustrate the relative contribution of each normalized feature (CiC_i). These components may be interpreted as effective measures of grittiness (PC1), strength (PC2), and toughness (PC3).

Analyzing DP Data Via Principal Component Analysis (PCA)

PCA was chosen due to its growing use within the field of classification [67]. To conduct the PCA we performed a two step normalization procedure. As {c1,c2,c3,c4}\{c_1, c_2, c_3, c_4\} were observed to be lognormally distributed, we first switched to the logarithmic values Ci=ln⁡(ci)C_i = \ln(c_i). Note that this implies C4=asinh⁡(N)C_4 = \operatorname{asinh}(N), a form often used to Gaussianize skewed, heavy-tailed data, particularly when zero-crossings are present. Next, each of the logarithmic CiC_i values was z-score normalized, as shown in Figure 2(c), which also reveals observable differences between the coordinate values for malignant and benign data. Prior to performing PCA, we confirmed the absence of multicollinearity [74] and thus the stability of the covariance structure necessary for the validity of this analysis (see Materials and Methods).

With the normalized data at hand, PCA was used to identify an alternative set of values which, through linear combinations of CiC_i, correspond to maximal variance in the data. For our PCA we utilize data from all 36 tests independently, as the goal is to capture the space in which tests vary. We find that the resulting first three principal components (PCs) (Figure 2(d-Figure 2e)) capture a majority (88.1%) of the variance in the dataset (PC1 [45.3%], PC2 [27.3%], PC3 [15.5%]).

Examining the weights of the different features for each principal component in Figure 2(d-Figure 2f) highlights major contributions and allows us to draw additional physical insights into thyroid nodule mechanics. Taking advantage of the logarithmic normalization we can accordingly write the following power law relationships

pc1=c10.4c40.6c20.3c30.6⟶effective grittiness,pc2=c10.6c20.8c30.1c40.1⟶effective strength,pc3=c10.4c30.8c40.3c20.2⟶effective toughness,(1)\begin{aligned} pc1 &= \frac{c_1^{0.4}c_4^{0.6}}{c_2^{0.3}c_3^{0.6}} \longrightarrow\text{effective } grittiness,\\ pc2 &= \frac{c_1^{0.6}c_2^{0.8}}{c_3^{0.1}c_4^{0.1}} \longrightarrow\text{effective } strength,\\ pc3 &= \frac{c_1^{0.4}c_3^{0.8}c_4^{0.3}}{c_2^{0.2}} \longrightarrow\text{effective } toughness, \tag*{(1)} \end{aligned}

where we have rounded the coefficients for compactness and have defined PCi=ln⁡(pci)\mathrm{PC}i = \ln(\mathrm{pc}i) (see Table S3 in Supplementary).

FeatureVIFPC1 LoadingPC2 LoadingPC3 LoadingD
C1C_11.250.43770.62620.42970.4359
C2C_21.21-0.34960.7610-0.17190.5552
C3C_31.27-0.5597-0.11820.81750.7069
C4C_41.450.6107-0.12150.34280.0437

Table S3. Multicollinearity evaluation of the features identified. Multicollinearity assessment (VIFs) of log-transformed features values CiC_i and PCA loadings for PC1-3 reconstruction and signed Euclidean distance from the plane in log-space.

Accordingly, pc1 contrasts contributions from critical insertion pressure and number of discrete rupture events against contributions from critical expansion pressure and critical expansion volume, with the number of discrete rupture events having the most significant role. Hence, we interpret pc1 as an effective measure of material grittiness. This aligns with the results of Sharma et al. [66], who have shown that “heterogeneity” (as determined by the number and size of rupture events) was a significant indicator for understanding different tissue types.

The second principal component, pc2, increases with the critical expansion and critical insertion pressures, and contains less dominant contributions from the critical expansion volume and the number of discrete rupture events. Since this component highlights the contributions from critical pressures we interpret pc2 as an effective measure of material strength.

Finally, the third principal component, pc3, increases with critical expansion volume, critical insertion pressure, and discrete rupture events, with a small contrasting contribution from the critical expansion pressure. We thus interpret this component as an effective measure of material toughness, reflecting the tissue’s capacity to accommodate deformation and dissipate energy before failure. The interpretations of the effective grittiness (pc1), strength (pc2), and toughness (pc3), provide physically intuitive axes along which to examine the mechanical distinctions between benign and malignant thyroid tissues, as well as among malignancy subtypes.

This is encouraging as it seems that quantitative measures analogous to the material properties which surgeons noted varied qualitatively across thyroid nodules capture the majority of variance in our data. Note here that a material stiffness, which surgeons often expressed as one of the most obvious factors between malignant and benign tumors, has yet to emerge. However, so far all analysis has been agnostic to the malignancy (or lack-thereof) of the nodules. In the following sections we will see that such a measure of stiffness does emerge as a deciding factor in nodule malignancy.

Nodule Subtype Data is Visually Distinguishable

Order reduction (i.e from four features for each DP test to three principal components) enables a more intuitive visualization of the data, which can aid in the identification of mechanically based separation between nodules that were later found to have different pathologies (see Table S1 and Table S2 in the Supplementary). In both graphs in Figure 3, each marker corresponds to one test and the marker shape represents the final pathology result: Papillary Thyroid Carcinoma (PTC), Oncocytic Thyroid Carcinoma (OTC), or Follicular Thyroid Carcinoma (FTC). In Figure 3(a), we consider two sub-groups, coloring benign pathologies in blue and malignant pathologies in red. To further assess whether separation exists, a kk-nearest neighbors (kk-NN) analysis with k=6k = 6 was applied to the 3D data and visualized as a projection in the PC1PC1–PC2PC2 plane. This visualization revealed a clear separation between malignant and benign nodules.

Figure 3. Visualization of separation between benign and malignant subtypes. Data from the PCA are displayed in the PC1-PC2 plane using the k-Nearest Neighbors (k-NN) algorithm with k=6k = 6 to color regions. (a) Regions that correspond to benign and malignant pathology results are visually separated. (b) Malignancy subtypes are also seen to form clusters.

To explore whether DP can differentiate between malignant subtypes themselves, Figure 3(b) presents the same kk-NN projection with each marker color corresponding to a distinct malignancy subtype (PTC, OTC, FTC). Natural clustering of malignancy subtypes is already observable in this visualization, suggesting that Digital Palpation has strong potential not only to distinguish malignant from benign tissue but also to discern between distinct malignancy subtypes in principal component space. Given our limited sample size, we avoid pairwise regressions between malignancy subtypes to prevent overfitting. Instead, this kk-NN visualization serves a critical illustrative purpose: clustering subtypes based on their mechanical properties is possible and justifies future work focused on increasing available mechanical data from thyroid nodules, which could further improve differentiation between subtypes

Statistical analysis of Malignancy Separation

A logistic regression model was chosen as the simplest tool to highlight the binary linear classification of whether a nodule was malignant or benign (Figure 3(a)). To perform this linear regression we again utilize the data from all 36 tests to capture the boundary between malignant and benign data.

Turning to a 3D representation of the DP data in Figure 4(a), we find that the grey plane, obtained via logistic regression, delineates the statistical boundary between these two groups. To quantify this separation, we computed the signed Euclidean distance from each test point to the logistic regression decision boundary plane. This metric provides a 1D measure of both significance (pp-value) and separability (Hedges’ gg) where positive distances lie on the malignant side of the boundary and negative distances on the benign side (see Materials and Methods for more information). As seen in Figure 4(b) the comparison between benign and malignant data using this approach yielded statistically significant separation with a pp-value of p=1×10−6p = 1 \times10^{-6} and g=1.94g = 1.94, confirming that the logistic regression designed to separate these two groups retains strong discriminatory power.

Visualization of PCA, logistic regression, and signed euclidean distance differentiation between benign and malignant thyroid nodules

Figure 4. Visualization of PCA, logistic regression, and signed euclidean distance differentiation between benign and malignant thyroid nodules. (a) Data from the PCA are displayed for the first three principal components, capturing 88.1% of the total variance. A solid grey plane denotes the logistic regression decision boundary separating benign (blue) and malignant (red) thyroid nodules. From the 3D PCA, we can directly translate each plot into a 1D representation using the signed Euclidean distance. (b) The 1D representation highlights that the benign versus malignant separation shows clear distinguishment with their mean ⟨μ⟩\langle\mu\rangle and 95% confidence interval. Additionally, using the same logistic regression applied for the benign versus malignant separation, we see distinguishment between benign and malignancy subtypes as well.

Additionally, we assess averaged nodule-level values to verify that the observed trends are not an artifact of intra-nodule dependence among repeated measurements. For this, we average the signed Euclidean distance of all 10 benign and malignant nodule values. Then we compare the five averaged benign nodules against the five averaged malignant nodules (visualization of the nodule-level data may be seen in Figure S2 in Supplementary). A two sample t-test on nodule-level mean signed distances yielded p=9×10−4p = 9 \times10^{-4} with a Hedges’ g=2.90g = 2.90 and 95% confidence bounds of [1.08, 4.66] calculated using a non-central t test. Note that even the lower bound of this Hedges’ gg estimate is greater than 0.8. This pp-value indicates that the separation between benign and malignant groups is statistically significant, while the Hedges’ gg indicates that the magnitude of this separation is substantial.

Per nodule signed Euclidean distances for benign and malignant nodules

Figure S2. Per nodule data collected during Digital Palpation. Signed euclidean distances for each nodule, corresponding with Table S1 and Table S2, with average values per nodule (and their equivalent 95% confidence limits) shown.

To account for multiple measurements being collected from the same nodule, we used a permutation-based statistical test that treated each nodule as a single independent unit of analysis [75]. This test asks, “If the true diagnostic label were unrelated to the measured features, how often would a random reassignment of nodules to diagnostic categories produce a separation as strong as our original assignment?” The permutation test yields a pp-value of p=2/252=8×10−3p = 2/252 = 8 \times10^{-3}. This is statistically significant and represents the strongest possible evidence achievable given our sample of 10 nodules. This confirms that our findings are robust to the nested structure of the data (multiple samples per nodule) and that the observed separation reflects genuine biological differences between benign and malignant nodule types, rather than chance clustering of individual measurements.

A Stiffness Metric Emerges from the Multivariate Analysis

The logistic regression decision boundary separating benign from malignant nodules defines a plane in the space of the first three principal components. Specifically, the signed Euclidean distance (DD) from any point to this plane in log space provides a scalar measure of malignancy risk, with positive values lying on the malignant side of the boundary. Because each principal component PCiPC_i is itself a linear combination of the original normalized features CiC_i, this distance can be expressed directly in terms of the measured mechanical quantities as

d=C10.4C20.6C30.7→effective ∗stiffness∗,(2)d = \frac{C_1^{0.4} C_2^{0.6}}{C_3^{0.7}} \to\text{effective } *stiffness*, \tag*{(2)}

where we have rounded the coefficients for compactness and defined D=ln⁡(d)D = \ln(d). Note that the weight of C4C_4 is much smaller than that those of {C1,C2,C3}\{C_1, C_2, C_3\} and was thus rounded to 0. This expression reveals that the primary axis of discrimination in our data is an effective measure of material stiffness: a ratio of pressures (needle insertion and cavity expansion) to the corresponding rupture volume during cavity expansion. Samples with higher values of this metric, i.e. those that develop higher pressures relative to the volume at which failure occurs, are more likely to be malignant. Remarkably, this mechanically-intuitive quantity emerges naturally from the multivariate analysis, demonstrating that the underlying physical difference between benign and malignant nodules can be captured by a single scalar reflecting the tissue’s effective stiffness and directly aligning with clinical intuition where clinicians often report malignant nodules as stiffer than benign nodules. This finding not only validates the mechanical basis of our approach but also provides a simple, interpretable metric with clear translational potential.

Conclusion

In this proof of principle work we show that mechanical data obtained via digital palpation (DP) exhibits strong potential as a minimally invasive method to inform thyroid cancer diagnosis. A straightforward, low-order statistical analysis of data from 36 ex vivo tests in 10 nodules (from 6 resected human thyroids), reveals clear statistical separation between benign and malignant nodules (p≤0.05p \le0.05) with a large effect size (Hedges’ g≥0.8g \ge0.8). Interestingly, this analysis yields a three-dimensional space of interpretable quantities that map closely to commonly reported clinical descriptions (i.e., grittiness, strength, and toughness). Performing logistic regression between benign and malignant nodules in this 3D space, we find that differentiation occurs along a single interpretable scalar parameter representative of stiffness.

Beyond the benign–malignant distinction, preliminary clustering (via k-nearest neighbors) in PC space was observed for the malignancy subtypes: follicular, oncocytic, and papillary thyroid carcinoma. These findings motivate the further development of DP not only as a method for distinguishing benign from malignant thyroid nodules, but also as a potential approach for identifying malignancy subtypes for improved clinical treatment strategies.

It is important to note that these ex vivo results may not directly translate to in vivo populations. A larger dataset would enable not only stronger differentiation between subpopulations, but also the use of statistical clustering methods (such as Gaussian mixture models) to determine subtype boundaries. Notably, the translation of DP to in vivo can utilize ultrasound guidance which may improve the precision of the test, but may introduce a new set of challenges.

In conclusion, DP may offer clinicians a minimally invasive, cost-effective, and precise tool that provides real-time information on thyroid nodule malignancy. Additionally, DP may be an applicable method to characterize disease in other tissues like breast or liver, ultimately reducing the number of avoidable surgeries and improving patient quality of life.

References and Notes

Acknowledgments

B.U. wishes to dedicate this paper to his sisters, Kristina and Alyssa, who were sources of inspiration for this work. The authors acknowledge the Newton-Wellesley Hospital Mass General Brigham (MGB) Pathology Team for their assistance with this project. J.B. would like to acknowledge the National Science Foundation (NSF) Graduate Research Fellowship Program as his primary funding source. H.V. acknowledges the support of the Department of Defense (DOD) through the National Defense Science & Engineering Graduate (NDSEG) Fellowship Program. S.K.N. acknowledges the MIT Postdoctoral Fellowship Program for Engineering Excellence (PFPFEE). S.P. and E.K. would like to acknowledge the Ruane fund at Massachusetts General Hospital (MGH).

Competing interests: There are no competing interests to declare.

Data, code and materials availability: All data and code are publicly available at the Cohen Mechanics Group GitHub (cohen-mechanics-group).

Supplementary Materials for Distinguishing thyroid nodule malignancy using digital palpation

Brendan M. Unikewicz; Joseph E. Bonavia; Ekaterina L. Koelliker; Abigail Klein; Katie C. Spaeth; Hannah M. Varner; S. Kiana Naghibzadeh; Yoav Horev; Sareh Parangi; Tal Cohen

Materials and Methods

Thyroid sample procurement and handling

In the course of this study, we performed a total of 36 DP tests in 10 different nodules. The number of tests in each nodule ranged between 2 to 7, depending on the size and location of the nodule, to ensure sufficient distance between insertion locations and from the nodule edge.

We followed a standard procedural protocol for thyroid sample and measurement. Excised human thyroids samples were procured at Newton-Wellesley Hospital in accordance with the IRB protocol 2008P001466. All methods were carried out in accordance with the IRB guidelines and regulations. Once the thyroidectomy was completed, the thyroid specimen was transferred to the on-site pathology room where the MIT team was prepared to perform DP prior to pathology evaluation. First, nodules were identified in consultation by the clinical team which included a pathologist. Initial and final diagnoses, from the pre-operative ultrasound, nodule size characteristics, and the final pathology diagnoses, were kept in a secured database for reference. DP test were performed immediately after surgical removal and testing occurred over the course of approximately one hour, depending on the number of nodules tested. Once testing was complete, the thyroid was then placed in formalin.

Experimental setup

Prior to a thyroid being received for testing, the experimental apparatus was prepared, on-site, at the Newton-Wellesley Hospital pathology room. The experimental apparatus was comprised of three distinct modules, each with their own Universal Serial Bus (USB) connections into a laptop computer: 1) the pressure sensing module, 2) the insertion module, and lastly, 3) the injection module (See Figure S1 in the Supplementary).

Digital Palpation experimental setup

Figure S1. Digital Palpation experimental setup. A thyroid specimen (1) containing nodules suspected of malignancy is positioned atop a vertically actuated translation stage (2) immediately after being resected from surgery. The stage raises the thyroid (1) into a fixed beveled needle (3), which penetrates the nodule. Following insertion, the cavity expansion protocol is initiated by displacing a syringe plunger (4) downward, injecting fluid at a controlled rate, via a high precision microinjector (5), into the tissue to form a fluid cavity.

Pressure sensing module

This module is comprised of a luer-tipped borosilicate glass 10 μL10\,\mu\mathrm{L} syringe (Hamilton 10 μL10\,\mu\mathrm{L} Syringe) filled with Phosphate-buffered saline (PBS) at a 9:1 ratio with distilled water, directly connected to a luer-locking pressure sensor (PRESS-S-0000), capped with a standard 25 Gauge (25G) hypodermic needle that is 25.4mm in length. Error estimates for the pressure sensor, depending on the recorded value, had a specification of ±2%\pm2\% for pressures recorded within 0 to 41.4kPa (0 to 6psi), ±3%\pm3\% for 41.4 to 206.7kPa (6 to 30psi), and ±3%\pm3\% for 0 to -48.3kPa (0 to -7psi). Analog to Digital Conversion (ADC) from the pressure sensor was performed via a 24-Bit ADC (Adafruit NAU7802). The ADC was interrogated over Inter-integrated circuit (I2C) protocols via a micro-controller USB key (Adafruit Trinkey QT2040) connected via a Qwiic connection. The micro-controller USB key was then plugged into a laptop computer (Lenovo Thinkpad X1 Carbon) and supplied a nominal 5.00VDC. The micro-controller USB key, then regulates the 5.00VDC to 3.30VDC for the 24-Bit ADC. The 24-Bit ADC then regulates the 3.30VDC to 3.00VDC to power the pressure sensor. The pressure sensor’s resolution was 0.2584mV/V/psi at a supply voltage of 3.00VDC leaving our effective system resolution at 0.7752mV/psi (0.1125mV/kPa). The micro-controller USB key was flashed with firmware to install the CircuitPython bootloader. The firmware file can be found at the cohen-mechanics-group GitHub under, u2if_trinkey_qt2040_rp2040.uf2. Interfacing with the pressure sensing module (and insertion and injection modules) is the Visual Studio (VS) Code Integrated Development Environment (IDE). Specifically, within the VS Code IDE, Python 3.7.0 was used for all hardware communication and data acquisition across all modules.

Insertion module

This model is comprised of a linear stage actuator (RATTMMOTOR EBX1605), for raising the thyroid specimen into the needle during testing. Mounted to the linear stage actuator is the translation stage which was printed out of Polylactic acid (PLA) plastic filament (Bambu Lab PLA Basic). This linear stage assembly was directly powered by a stepper motor controller (Polulu Tic 36v4) and the stepper motor controller was powered via a 8.5A 24V AC/DC power supply connected to standard wall outlet (∼\sim110VAC). Communications with the stepper motor controller were handled over serial via a USB Type-A to Micro-B cable through the VS Code IDE with the same control software interfacing with the pressure sensing and injection modules.

Injection module

This module is comprised of a syringe microinjector (World Precision Instruments UMP3T) and the microinjector controller (World Precision Instruments MICROTOUCH 2T). Specifically, the syringe microinjector is connected to the microinjector controller via serial cable to the rear panel of the microinjector controller. Next, the microinjector controller is powered by a traditional wall outlet (∼110VAC\sim110\mathrm{VAC}) and additionally connected to the host laptop computer via a USB A-to-B cable. On the microinjector controller, remote access is enabled, such that the host laptop computer can remotely access and program the microcontroller for controlled injections via the VS Code IDE through Python 3.7.0 for concurrent pressure measurement with the pressure sensing module. Once setup, the pressure sensing module, e.g. the glass syringe is loaded into the syringe microinjector and the experimental apparatus is setup.

Thyroid nodule testing protocol

The testing protocol strictly followed the five following steps:

DP device preparation and calibration test

Prior to receiving thyroid specimens, a water calibration test was performed. For this, with the glass syringe completely filled with PBS solution, an injection procedure was performed into a 1oz disposable plastic cup to measure the pressure drop along the needle. For this, the injection protocol was identical to the DP testing protocol with a liquid transfer rate of 0.6 μL/s0.6\,\mu\mathrm{L}/\mathrm{s} up to a fluid volume of 9.2 μL9.2\,\mu\mathrm{L}. By inspecting the obtained pressure-volume calibration curve for expected values we confirmed there was no presence of air bubbles and that all components are operating as intended. Once the water calibration was completed, and the system was verified to have no air entrapment, the glass syringe was refilled with the PBS solution and thyroid tissue specimen testing could commence.

Thyroid assessment and preparation

Upon receiving a resected thyroid, post-surgery, the thyroid was rested on a glass petri dish and placed on top of the linear stage assembly. Nodules were located with the assistance of the clinical team, and the number and choice of nodules tested were determined by the time available prior to tissue fixation.

Insertion procedure

With the device and thyroid ready for testing, the computer was used to initialize the testing protocol, beginning with the insertion procedure. First, the linear stage assembly was translated upwards until the 25G hypodermic needle made first contact with the surface of the thyroid specimen, positioned directly over top the thyroid nodule. Needle insertion was then performed at a controlled insertion rate of 12.50mm/s12.50\mathrm{mm}/\mathrm{s}, by elevating the translation stage while concurrently recording the pressure. Insertion stopped upon arriving at the desired depth (see insertion depths in Table S1 and Table S2 of the Supplementary). Approximately one minute after the insertion was complete, the expansion procedure was performed.

Expansion procedure

Following insertion, the injection protocol was performed. Specifically, fluid was transferred via the syringe microinjector at a rate of 0.6μL/s0.6\mathrm{\mu L}/\mathrm{s} for a total injected fluid volume of 9.2μL9.2\mathrm{\mu L}.

Test completion and device sterilization

Upon completion of the DP test, the hypodermic needle was removed from the thyroid nodule via translating the stage downwards. The hypodermic needle was then placed in a 10% bleach solution, and the translation stage was moved up and down, while periodically withdrawing and expelling 10% bleach solution (∼0.5μL\sim0.5\mathrm{\mu L}) to cleanse the outer and inner portions of the hypodermic needle. Afterwards, the glass syringe was fully purged of any remaining liquid. Once the glass syringe assembly was fully emptied into the PBS solution, the glass syringe assembly was refilled with PBS, a water calibration performed once again to verify all components were operating as intended without the introduction of air bubbles, and thyroid nodule testing could commence once again. To test a same thyroid nodule, the thyroid specimen would be repositioned and reoriented such that needle insertion sites into the thyroid nodule were unique.

Extraction of characteristic features from DP data

From the insertion and injection curves, four key features were extracted via MATLAB (see Figure 1 and Table S5). First, the cavity expansion curve was calibrated by subtracting the pressure drop along the needle. This calibration accounts for pressure losses between the sensor and the needle tip as fluid is expelled into the thyroid specimen. The peak insertion pressure, c1c_1, is defined as the maximum recorded pressure during insertion. The peak pressure during cavity expansion, denoted c2c_2, was determined using a prominence-based peak detection algorithm (MATLAB findpeaks command). To distinguish physiologically meaningful rupture events from signal noise, only peaks with a prominence and height exceeding a threshold of 120\frac{1}{20} of the mean pressure were retained. Among all detected peaks meeting these criteria, the peak with the largest amplitude was selected as c2c_2, and the corresponding fluid volume at that point was recorded as c3c_3. The total number of detected peaks satisfying the criteria, denoted NN, was transformed as c4=N+1+N2c_4 = N + \sqrt{1 + N^2} to regularize the discrete count and accommodate true zeros. If no peaks met the detection criteria, NN was set to zero, and c2c_2 and c3c_3 were defined as the pressure and volume at 99% of the total expansion duration, ensuring that end-of-test artifacts did not bias the measurement.

#Final PathologyC1C_1C2C_2C3C_3C4C_4
1Benign-0.14481.42132.27553.5272
2Benign-0.55211.46112.29481.8184
3Benign0.12151.05802.75591.8184
4Benign0.40421.88993.04340.0000
5Benign-2.26940.83742.12813.3335
6Benign-0.88313.02723.04330.0000
7Benign-0.93842.81753.04340.0000
8Benign0.65542.13473.01831.8184
9Benign0.69831.77972.48793.5272
10Benign-1.52381.14531.79602.8934
11Benign0.88791.12041.47944.2487
12Benign-2.03702.38802.83501.8184
13Benign2.03952.86891.81541.4436
14Benign1.5092-0.54982.92674.4069

Table S5. Log-transformed features values CiC_i for benign nodule data. Final pathology results shown against the log-transformed feature values ascertained through Digital Palpation.

Principal Component Analysis (PCA)

To perform PCA we first compute variance inflation factors (VIFs) to assess multicollinearity among log transformed input features (see Table S3 in Supplementary). All VIF values were below 5.00, ensuring acceptable levels of multicollinearity [74] and supporting the stability of the covariance structure necessary for PCA. We note that in the feature normalization procedure described in the main text, we included an offset of 10−610^{-6} kPa to c1c_1 to ensure there was no zero-handling during the analysis of c1c_1.

Logistic Regressions and Signed Euclidean Distances

Logistic regression was employed as a probabilistic classifier using the log-transformed features projected into principal component space, with model parameters estimated via maximum likelihood. Distinct logistic regression models were fitted to address three tasks: (i) distinguishing benign from malignant samples using the first three principal components; (ii) a comparison between PTC and the combined OTC+FTC subtypes to assess global separability of PTC from other malignancies; and (iii) subtype-specific distinguishment comparing each malignant subtype pair (PTC vs. OTC, PTC vs. FTC, and OTC vs. FTC) to evaluate inter-subtype separability and significance. For each model, the full set of regression coefficients β0\beta_0 through β3\beta_3 are reported, enabling reconstruction of the three-dimensional decision plane defined by β0+β1PC1+β2PC2+β3PC3=0\beta_0+\beta_1\mathrm{PC}_1+\beta_2\mathrm{PC}_2+\beta_3\mathrm{PC}_3=0. From each model, signed Euclidean distances were computed as the orthogonal projection onto the axis normal to the decision plane: D=D=

β1PC1+β2PC2+β3PC3β12+β22+β32\frac{\beta_1\mathrm{PC}_1+\beta_2\mathrm{PC}_2+\beta_3\mathrm{PC}_3}{\sqrt{\beta_1^2+\beta_2^2+\beta_3^2}}

To quantify significance, two-sample tt-tests were performed and to understand separability, the Hedges’ gg was used to understand mean effect size. To provide readers with an intuitive visualization of subtype clustering independent of parametric assumptions, a k-Nearest Neighbors analysis (k=6k=6) was performed in the PC1\mathrm{PC}_1-PC2\mathrm{PC}_2 plane; this visualization is included for interpretive purposes only and was not used for statistical or separability analysis.

Supplementary: Tables and Figures

Comparisonpp-valueHedges’ ggβ0\beta_0β1\beta_1 (PC1)β2\beta_2 (PC2)β3\beta_3 (PC3)
---β0\beta_0β1\beta_1 (PC1)β2\beta_2 (PC2)β3\beta_3 (PC3)
Benign vs Malignant0.0000011.940.87430.86041.846-1.179
Nodule Averaged: Benign vs Malignant0.00092.90————

Table S4. Measures of statistical significance as well as values for 3D plane reconstruction. Signed distance pp- and gg-values for benign versus malignant nodules as well as logistic regression coefficients for the 3D decision boundary.

#Final PathologyC1C_1C2C_2C3C_3C4C_4
1Malignant0.31782.86402.15360.8814
2Malignant-0.16902.99542.15851.8184
3Malignant-2.86024.00632.44071.4436
4Malignant-0.04022.11841.90390.8814
5Malignant-0.33901.41541.76281.8184
6Malignant2.17071.21672.87714.3309
7Malignant0.91382.62202.05312.0947
8Malignant2.17652.28471.53091.8184
9Malignant2.96181.63571.70912.0947
10Malignant3.09493.02421.62382.4918
11Malignant1.76361.24071.44183.9124
12Malignant3.12495.26642.92422.4918
13Malignant1.44763.12591.57684.5851
14Malignant1.44313.67701.74394.0257
15Malignant2.05104.15121.82132.6441
16Malignant1.38550.95361.43832.7765
17Malignant1.28950.39321.48293.1798
18Malignant1.21451.58581.60513.2596
19Malignant1.19111.43491.67282.0947
20Malignant1.35521.56271.57462.6441
21Malignant0.85400.72221.43783.4667
22Malignant0.73331.40531.44843.4667

Table S6. Log-transformed features values CiC_i for malignant nodule data. Final pathology results shown against the log-transformed feature values ascertained through Digital Palpation.

References

  1. [1]Y. Li, et al., Prevalence of thyroid nodules in China: a health examination cohort-based study. Frontiers in endocrinology 12, 676144 (2021).DOI
  2. [2]S. Tamhane, H. Gharib, Thyroid nodule update on diagnosis and management. Clinical diabetes and endocrinology 2 (1), 17 (2016).DOI
  3. [3]P. D. Gutman, M. Henry, Fine needle aspiration cytology of the thyroid. Clinics in laboratory medicine 18 (3), 461–482 (1998).pubmed.ncbi.nlm.nih.gov/9742379
  4. [4]R. HaugenBryan, et al., 2015 American Thyroid Association management guidelines for adult patients with thyroid nodules and differentiated thyroid cancer: the American Thyroid Association guidelines task force on thyroid nodules and differentiated thyroid cancer. Thyroid (2016).
  5. [5]C. Durante, et al., The diagnosis and management of thyroid nodules: a review. Jama 319 (9), 914–924 (2018).DOI
  6. [6]National Cancer Institute, SEER Stat Fact Sheets: Thyroid Cancer, https://seer.cancer.gov/statfacts/html/thyro.html (2024), last accessed December 19, 2024.
  7. [7]S. Vaccarella, et al., Global patterns and trends in incidence and mortality of thyroid cancer in children and adolescents: a population-based study. The Lancet Diabetes & Endocrinology 9 (3) (2021-03).DOI
  8. [8]K. L. Yan, et al., Rising incidence and incidence-based mortality of thyroid cancer in California, 2000-2017. The Journal of Clinical Endocrinology & Metabolism 105 (6), 1770–1777 (2020).DOI
  9. [9]M. Luster, L. H. Duntas, L. Wartofsky, The Thyroid and Its Diseases (Springer) (2019).
  10. [10]M. R. Haymart, et al., Variation in the management of thyroid cancer. The Journal of Clinical Endocrinology & Metabolism 98 (5), 2001–2008 (2013).DOI
  11. [11]C. D. Seib, J. A. Sosa, Evolving understanding of the epidemiology of thyroid cancer. Endocrinology and Metabolism Clinics 48 (1), 23–35 (2019).DOI
  12. [12]International Agency for Research on Cancer (IARC), Cancer Today — Pie visualization (All cancers, both sexes, ages 0–17) (2024), https://gco.iarc.who.int/today/en/dataviz/pie?mode=cancer&group_populations=1&sexes=2&age_end=17.
  13. [13]T. M. Vaghaiwalla, E. M. Henriksen, J. Chang, C. Saghira, J. I. Lew, Reassessing false-negative rate and size cutoff for papillary thyroid cancer with fine-needle aspiration in thyroid nodules. Surgery 186, 109577 (2025).DOI
  14. [14]S. J. Nam, et al., Large (≥ 3cm) thyroid nodules with benign cytology: Can Thyroid Imaging Reporting and Data System (TIRADS) help predict false-negative cytology? PLoS One 12 (10), e0186242 (2017).DOI
  15. [15]E. D. Rossi, A. J. Adeniran, W. C. Faquin, Pitfalls in thyroid cytopathology. Surgical pathology clinics 12 (4), 865 (2019).DOI
  16. [16]M. Bongiovanni, A. Spitale, W. C. Faquin, L. Mazzucchelli, Z. W. Baloch, The Bethesda system for reporting thyroid cytopathology: a meta-analysis. Acta cytologica 56 (4), 333–339 (2012).DOI
  17. [17]B. K. Richmond, et al., False-negative results with the Bethesda System of reporting thyroid cytopathology: predictors of malignancy in thyroid nodules classified as benign by cytopathologic evaluation. The American surgeon 80 (8), 811–816 (2014).pubmed.ncbi.nlm.nih.gov/25105404
  18. [18]R. Kulstad, Do all thyroid nodules > 4 cm need to be removed? An evaluation of thyroid fine-needle aspiration biopsy in large thyroid nodules. Endocrine Practice 22 (7), 791–798 (2016).DOI
  19. [19]C. M. Kitahara, J. A. Sosa, The changing incidence of thyroid cancer. Nature Reviews Endocrinology 12 (11), 646–653 (2016).DOI
  20. [20]C. M. Kitahara, J. A. Sosa, Understanding the ever-changing incidence of thyroid cancer. Nature Reviews Endocrinology 16 (11), 617–618 (2020).DOI
  21. [21]C. Saint-Martin, et al., Overdiagnosis of thyroid cancer in the Marne and Ardennes Departments of France from 1975 to 2014, in Annales d’Endocrinologie (Elsevier), vol. 78 (2017), pp. 27–32.
  22. [22]H. S. Ahn, H. J. Kim, H. G. Welch, et al., Korea’s thyroid-cancer “epidemic”—screening and overdiagnosis. N Engl J Med 371 (19), 1765–1767 (2014).DOI
  23. [23]L. Davies, H. G. Welch, Current thyroid cancer trends in the United States. JAMA otolaryngology–head & neck surgery 140 (4), 317–322 (2014).DOI
  24. [24]S. Alshaikh, Z. Harb, E. Aljufairi, S. A. Almahari, Classification of thyroid fine-needle aspiration cytology into Bethesda categories: An institutional experience and review of the literature. Cytojournal 15 (2018).DOI
  25. [25]A. D. Workman, et al., Limitations of preoperative cytology for medullary thyroid cancer: proposal for improved preoperative diagnosis for optimal initial medullary thyroid carcinoma specific surgery. Head & Neck 43 (3), 920–927 (2021).DOI
  26. [26]P. Trimboli, et al., Detection rate of FNA cytology in medullary thyroid carcinoma: a meta-analysis. Clinical Endocrinology 82 (2), 280–285 (2015).DOI
  27. [27]A. S. Ho, et al., Malignancy rate in thyroid nodules classified as Bethesda category III (AUS/FLUS). Thyroid 24 (5), 832–839 (2014).DOI
  28. [28]M. Endo, et al., Features of cytologically indeterminate molecularly benign nodules treated with surgery. The Journal of Clinical Endocrinology & Metabolism 105 (11), e3971–e3980 (2020).DOI
  29. [29]B. R. Haugen, et al., 2015 American Thyroid Association management guidelines for adult patients with thyroid nodules and differentiated thyroid cancer: the American Thyroid Association guidelines task force on thyroid nodules and differentiated thyroid cancer. Thyroid 26 (1), 1–133 (2016).DOI
  30. [30]M. N. Mongelli, et al., Financial burden and quality of life among thyroid cancer survivors. Surgery 167 (3), 631–637 (2020).DOI
  31. [31]C. C. Lubitz, J. A. Sosa, The changing landscape of papillary thyroid cancer: epidemiology, management, and the implications for patients. Cancer 122 (24), 3754–3759 (2016).DOI
  32. [32]M. K. Applewhite, et al., Quality of life in thyroid cancer is similar to that of other cancers with worse survival. World journal of surgery 40 (3), 551–561 (2016).DOI
  33. [33]B. Aschebrook-Kilfoy, et al., Risk factors for decreased quality of life in thyroid cancer survivors: initial findings from the North American Thyroid Cancer Survivorship Study. Thyroid 25 (12), 1313–1321 (2015).
  34. [34]M. R. Burge, T.-M. Zeise, M. W. Johnsen, M. J. Conway, C. R. Qualls, Risks of complication following thyroidectomy. Journal of general internal medicine 13 (1), 24–31 (1998).pubmed.ncbi.nlm.nih.gov/9462491
  35. [35]J. Liu, et al., Risk factors for post-thyroidectomy haemorrhage: a meta-analysis. European journal of endocrinology 176 (5), 591–602 (2017).DOI
  36. [36]J. Sancho, et al., Risk factors for transient vocal cord palsy after thyroidectomy. Journal of British Surgery 95 (8), 961–967 (2008).DOI
  37. [37]N. Christou, M. Mathonnet, Complications after total thyroidectomy. Journal of visceral surgery 150 (4), 249–256 (2013).DOI
  38. [38]A. Tran, S. Hyer, I. Rafi, O. Okosieme, Thyroid hormone replacement in the preconception period and pregnancy. The British Journal of General Practice 69 (683), 282 (2019).DOI
  39. [39]J. F. Krane, et al., The Afirma Xpression Atlas for thyroid nodules and thyroid cancer metastases: Insights to inform clinical decision-making from a fine-needle aspiration sample. Cancer cytopathology 128 (7), 452–459 (2020).DOI
  40. [40]R. C. Jug, M. B. Datto, X. S. Jiang, Molecular testing for indeterminate thyroid nodules: performance of the Afirma gene expression classifier and ThyroSeq panel. Cancer Cytopathology 126 (7), 471–480 (2018).DOI
  41. [41]N. P. Ohori, et al., Benign call rate and molecular test result distribution of ThyroSeq v3. Cancer cytopathology 127 (3), 161–168 (2019).DOI
  42. [42]T. Chen, et al., The role of the ThyroSeq v3 molecular test in the surgical management of thyroid nodules in the Canadian public health care setting. Thyroid 30 (9), 1280–1287 (2020).DOI
  43. [43]J. Patel, J. Klopper, E. E. Cotrill, Molecular diagnostics in the evaluation of thyroid nodules: current use and prospective opportunities. Frontiers in endocrinology 14, 1101410 (2023).DOI
  44. [44]M. J. Livhits, et al., Effectiveness of molecular testing techniques for diagnosis of indeterminate thyroid nodules: a randomized clinical trial. JAMA oncology 7 (1), 70–77 (2021).DOI
  45. [45]C. DiGennaro, et al., Assessing Bias and Limitations of Clinical Validation Studies of Molecular Diagnostic Tests for Indeterminate Thyroid Nodules: Systematic Review and Meta-Analysis. Thyroid (2022).
  46. [46]C. Wong, X. Liu, B. Lang, Cost-effectiveness of fine-needle aspiration cytology (FNAC) and watchful observation for incidental thyroid nodules. Journal of endocrinological investigation 43 (11), 1645–1654 (2020).DOI
  47. [47]B. Aschebrook-Kilfoy, et al., The Clinical and Economic Burden of a Sustained Increase in Thyroid Cancer IncidenceIncreasing Thyroid Cancer Burden. Cancer epidemiology, biomarkers & prevention 22 (7), 1252–1259 (2013).
  48. [48]D. L. Steward, et al., Performance of a multigene genomic classifier in thyroid nodules with indeterminate cytology: a prospective blinded multicenter study. JAMA oncology 5 (2), 204–212 (2019).DOI
  49. [49]K. N. Patel, et al., The American Association of Endocrine Surgeons guidelines for the definitive surgical management of thyroid disease in adults. Annals of surgery 271 (3), e21–e93 (2020).DOI
  50. [50]T. M. Khan, M. A. Zeiger, Thyroid nodule molecular testing: is it ready for prime time? Frontiers in Endocrinology 11, 590128 (2020).DOI
  51. [51]P. N. Wells, H.-D. Liang, Medical ultrasound: imaging of soft tissue strain and elasticity. Journal of the Royal Society Interface 8 (64), 1521–1549 (2011).DOI
  52. [52]N. Nitta, M. Yamakawa, H. Hachiya, T. Shiina, A review of physical and engineering factors potentially affecting shear wave elastography. Journal of Medical Ultrasonics 48 (4), 403–414 (2021).
  53. [53]T. Anbarasan, C. Wei, J. C. Bamber, R. G. Barr, G. Nabi, Characterisation of prostate lesions using transrectal shear wave elastography (SWE) ultrasound imaging: A systematic review. Cancers 13 (1), 122 (2021).DOI
  54. [54]Y. J. Yoo, et al., Computer-aided diagnosis of thyroid nodules via ultrasonography: initial clinical experience. Korean journal of radiology 19 (4), 665–672 (2018).DOI
  55. [55]H. J. Moon, et al., Diagnostic performance of gray-scale US and elastography in solid thyroid nodules. Radiology 262 (3), 1002–1013 (2012).DOI
  56. [56]R. M. Sigrist, J. Liau, A. El Kaffas, M. C. Chammas, J. K. Willmann, Ultrasound elastography: review of techniques and clinical applications. Theranostics 7 (5), 1303 (2017).DOI
  57. [57]E. A. Selvaraj, et al., Diagnostic accuracy of elastography and magnetic resonance imaging in patients with NAFLD: a systematic review and meta-analysis. Journal of hepatology 75 (4), 770–785 (2021).
  58. [58]D. Toro-Tobon, et al., Artificial intelligence in thyroidology: a narrative review of the current applications, associated challenges, and future directions. Thyroid 33 (8), 903–917 (2023).
  59. [59]M. M. A. Alabrak, et al., Artificial intelligence role in subclassifying cytology of thyroid follicular neoplasm. Asian Pacific Journal of Cancer Prevention: APJCP 24 (4), 1379 (2023).DOI
  60. [60]J. Hoang, Thyroid nodules and evaluation of thyroid cancer risk. Australasian journal of ultrasound in medicine 13 (4), 33 (2015).
  61. [61]G. Azizi, et al., Performance of elastography for the evaluation of thyroid nodules: a prospective study. Thyroid 23 (6), 734–740 (2013).DOI
  62. [62]J. Luo, et al., Risk of malignancy in thyroid nodules: predictive value of puncture feeling of grittiness in the process of fine-needle aspiration. Scientific Reports 7 (1), 13109 (2017).DOI
  63. [63]F. Rosenzweig, R. Feruglio, T. Marin, Palpation and Its Learning: A Professionalization Approach to Acquiring a Complex Skill. International Journal of Osteopathic Medicine p. 100792 (2025).DOI
  64. [64]P. Hannequin, J. C. Liehn, B. Maes, M. J. Delisle, Multivariate analysis in solitary cold thyroid nodules for the diagnosis of malignancy. European Journal of Cancer and Clinical Oncology 24 (5), 881–888 (1988).pubmed.ncbi.nlm.nih.gov/3169093
  65. [65]J. Wu, Y. Li, M. Zhang, Clinical value of FNA puncture feeling in the diagnosis of non-diagnostic and indeterminate thyroid nodules. Frontiers in Endocrinology 13, 1022438 (2022).DOI
  66. [66]S. Sharma, R. Aguilera, J. Rao, J. K. Gimzewski, Piezoelectric needle sensor reveals mechanical heterogeneity in human thyroid tissue lesions. Scientific reports 9 (1), 9282 (2019).DOI
  67. [67]M. Greenacre, et al., Principal component analysis. Nature Reviews Methods Primers 2 (1), 100 (2022).
  68. [68]S. K. Naghibzadeh, H. Rudykh, B. M. Unikewicz, T. Cohen, A Novel Approach to Measure Needle Insertion Response and the Overlooked Impact of Insertion on Injection Outcomes. Soft Matter (2026).
  69. [69]S. Raayai-Ardakani, Z. Chen, D. R. Earl, T. Cohen, Volume-controlled cavity expansion for probing of local elastic properties in soft materials. Soft matter 15 (3), 381–392 (2019).
  70. [70]S. Chockalingam, C. Roth, T. Henzel, T. Cohen, Probing local nonlinear viscoelastic properties in soft materials. Journal of the Mechanics and Physics of Solids 146, 104172 (2021).DOI
  71. [71]A. S. Mijailovic, et al., Localized characterization of brain tissue mechanical properties by needle induced cavitation rheology and volume controlled cavity expansion. Journal of the Mechanical Behavior of Biomedical Materials 114, 104168 (2021).DOI
  72. [72]H. Varner, G. P. Sugerman, M. K. Rausch, T. Cohen, Elasticity of whole blood clots measured via Volume Controlled Cavity Expansion. Journal of the Mechanical Behavior of Biomedical Materials 143, 105901 (2023).DOI
  73. [73]H. M. Varner, S. Naghibzadeh, K. Spaeth, A. Klein, T. Cohen, Evolving properties of biological materials captured via needle-based cavity expansion method. Experimental Mechanics 65 (1), 141–155 (2025).DOI
  74. [74]J. H. Kim, Multicollinearity and misleading statistical results. Korean journal of anesthesiology 72 (6), 558–569 (2019).DOI
  75. [75]E. L. Lehmann, J. P. Romano, Testing statistical hypotheses (Springer) (2005).

Paper details

vlnc:W_7y5xmf55
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

Not recorded for this paper

View statement
Looking for related papers…
Contents