🌏
Premier International Publisher
Serving Researchers Since 2012

THYROX: A Unified Explainable Deep Learning System for Thyroid Nodule Classification and Lymph Node Metastasis Assessment in Ultrasound

DOI : 10.5281/zenodo.24397658
Download Full-Text PDF Cite this Publication

Text Only Version

THYROX: A Unified Explainable Deep Learning System for Thyroid Nodule Classification and Lymph Node Metastasis Assessment in Ultrasound

Tanish P Jadhav

Dept of CSE(AI&ML)

Don Bosco Institiute of Technology Bengaluru, India

Sarvesh Patil

Dept of CSE(AI&ML)

Don Bosco Institute of Technology Bengaluru, India

Prof. Padmavathi B

Dept of CSE(AI&ML)

Don Bosco Institute of Technology Bengaluru, India

Abstract – Thyroid nodule is common but rarely malignant, and ultrasound interpretation remains subject to substantial inter-observer variability. In papillary thyroid carcinoma (PTC), staging additionally requires assessment of the cervical lymph node basin, which is frequently under-detected preoperatively. Existing deep learning systems treat the nodule and the nodal basin separately, rarely quantify their explanation and are seldom validated on public data. We present THYROX, a unified decision support system performing nodule malignancy classification, lesion localization, risk stratification and cervical nodal assessment in one explainable pipeline. Two public datasets were used:TN5000(5,000 biopsy confirmed B- mode images with bounding box annotations) and LymphUs (338 PTC patients, two centers). Four ImageNet pretrained backbones ResNet50, DenseNet121, EfficientNetV2-S and ConvNeXt-Tiny were fine-tuned under an identical protocol and combined by soft vote ensembling with three view test time augmentation, oversampling only the training split. For the nodal task a DenseNet121 was warm started from the thyroid checkpoint rather than ImageNet, testing whether an in-domain ultrasound representation transfers across anatomical sites under label scarcity. On the TN5000 test set ( = ) the ensemble achieved 91.87% accuracy (95% CI 89.69-93.62),

95.34% sensitivity, 83.18% specificity and ROC-AUC 0.9675. on LymphUs ( = ) the warm-started model reached 88.24% accuracy with 100% specify from only 233 trained images. Three backbones performed equivalently (AUC range 0.0021) while ConvNeXt-Tiny failed to converge, and single-box regression proved inadequate for localization (mean IoU 0.286); both negative results are reported.

Keywords: thyroid nodule, cervical lymph node metastasis, ultrasound, deep learning, transfer learning, explainable AI, Grad-Cam, ensemble learning

  1. INTRODUCTION

    Thyroid cancer is the most common endocrine malignancy, with an estimate over 820,000 new cases and 47,000 deaths worldwide in 2022 [1]. Recorded incidence has risen steeply for four decades while mortality has remained flat a divergence attribute to over detection of small, indolent papillary carcinomas rather than a true rise in disease burden [1]. The clinical problem is therefore not detection but discrimination.

    Ultrasonography is the first-line modality. Structured re- porting, principally ACR TI-RADS [25], translates sonographic features into risk categories and fine needle aspiration (FNA) thresholds, reducing biopsy rates while improving accuracy [21]. The features themselves, however

    are assessed with only moderate reproducibility: pooled meta- analysis report inter-rater agreement of 0.51 0.51 [13]. Downstream, FNA cytology returns an indeterminate result in a substantial fraction of sampled nodules, driving diagnostic surgery in patients with benign disease [24].

    A second problem concerns disease extent. Cervical lymph node metastasis is common in PTC and materially changes surgical planning [27], yet preoperative ultrasound detection is far less sensitive: a meta-analysis of 27 studies covering more than 23,000 patients reports pooled radiologist sensitivity of 0.51 against 0.80 for AI models [6].

    Deep learning has addressed the first problem extensively [19], [9], [3]; nodal prediction has followed a smaller trajectory [5], [6]. Three limitations recur. First, the nodule and the nodal basin are treated separately despite being assessed in the same examination for the same decision. Second, explanations are displayed but not evaluated, although benchmarking in other modalities shows saliency methods localising markedly worse than human annotation [14]. Third, evaluation often uses small private cohorts, sometimes with resampling applied before partitioning, which inflates reported accuracy.

    This work addresses all three. Our contributions are:

    1. A unified two-task pipeline covering nodules malignancy, localisation and risk stratification together with a prallel cervical nodal module, mirroring clinical PTC staging.

    2. Cross anatomical in domain transfer: the nodal classifier is warm started from a thyroid ultrasound that speckle statistics and margin priors are shared across ultrasound anatomies.

    3. A four architechture benchmark with ensembling under one shared regularisation protocol, isolating back-bone choice from training recipe.

    4. Quanttitaive explainability, evaluating Grad-CAM against TN5000s ground-truth boxes rather than pre senting it illustratively.

    5. Public data, reported negative results and released code.

  2. RELATED WORK

    1. Noduel classification, nodal metastais and transfer

      Multicohort convolutional studies established that networks approach expert performance and that the endpoint is better framed as a biopsy-management decision [19], and that assistance raises junior readers toward senior level accuracy prospectively [9]; the field has since moved toward transformer and hybrid designs [3], with Rashed et al. [4] the closest analogue to this work. Nodal prediction is younger progressing from an initial pilot [22] to explainable multi- modal learning [5]; meta-analysis identifies the same gap throughout most studies are single-centre and few release data [6]. On initialisation, Raghu et al. [20] showed that much of ImageNet transfers benefit derives from weight scaling rather than feature reuse, and benchmarks confirm in-domain pretraining generally wins under label scarcity [12]. Most relevant here, USFM [10] demonstrated that ultrasound representations generalise across organs. Grad-CAM [17] is the standard explanation tool, yet across seven saliency methods on chest radiographs all localised significantly worse than a human benchmark [14], and quantitative protocols are rarely applied to thyroid ultrasound.

    2. Positioning against immediate atecedents

    Three studies from the direct antecedents of this work(Table I). Nugroho et al. [26] classify nodule margin from nine geometric descriptors with SMOTE balancing, reporting 94.79% accuracy on 165 private images. We adopt the descriptor family but differ in endpoint, data provenance,

    yielding 338 images (175 benign, 163 malignant). Dataset properties and partitioning appear in Table II.

    Partitioning was stratified by class. Because the TN5000 release does bot expose patient identifiers, thyroid partitioning is performed at the image level; no image appears in more that one split, but patient level disjointness cannot be claimed. Nodal partitioning is patient-level and verified disjoint.

    1. Pre-processing and augmentation

      An input frame H×W×3 is bilinearly resized to 224×224 and normalized channel-wise as x(c)=(x(c)/255-µc)/c with the ImageNet statistics µ=(0.485, 0.456, 0.406), =(0.229, 0.224, 0.225). Augmentation (Table III) uses only ultrasound plausible tansforms, with random erasing acting as a cutout regularizer simulating acoustic shadowing. A heavier variant was used for the nodal task, where the training set is an order of magnitude smaller and oversampled duplicates must be rendered visually distinct at each epoch.

      Critically, the two branches take separate images: a thyroid image is not a valid input to the nodal classifier, and the interface reports not assessed for any branch whose input is absent.

    2. Localisation and geometric features

      Lesion localisation used a ResNet50 backbone with a four unit sigmoid head regressing the box in normalised coordinates b=(min / W, min / H, max / W, max / H) [0,1]4, trained with the Smooth-L1 objective

      and importantly in applying oversampling only after partitioning: in [18] SMOTE expands 165 images to 480 instance before the 80/20 split, so synthetic samples interpolated from test set originals enter training. Pimpalkar

      box

      4

      = (

      =1

      ), () = {

      1 2, || < 1

      2

      1

      || , otherwise, 2

      (1)

      and Gulame [27] fuse transverse and longitudinal views through two YOLOv5 detectors and an XGBoost head, reporting 99.64% accuracy on an 81 case test set; we adopt the

      Using AdamW at 1 × 10-4 for up to 20 epochs (patience 5) Localisation quality is measured by the Jaccard index

      multi-view and detector then classifier structure, while noting their component detectors reach mAP@0.5 of only 0.70-

      0.72. Dency Flora et al. [11] survey thyroid dysfunction

      IoU(, ) = ||

      ||

      (2)

      classification from laboratory attributes rather than malignancy from images; we cite it for motivation and exclude it from performance comparison.

  3. MATERIALS AND METHODS

    The architecture is shown in Fig. 1; this section follows it stage by stage.

    1. Datasets

      TN5000 [7] compromises 5,000 B-mode thyroid ultra- sound images 3,572 malignant and 1,428 benign labelled by FNA biopsy with pathological confirmation and annotated by

      Reported as the mean over the test set and as the fraction exceeding 0.5.

      Nine geometric margin descriptors follow the formulation of [18], computed from the largest external contour recovered by median filtering, CLAHE, Otsu thresholding and elliptical morphological closing and opening within the region of interest. Writing A and P for contour area and perimeter, Ah and Ph for those of the convex hull, and di = || pi – ||2 for the radial distance from the centroid to boundary point , the three most discriminative are

      = 4 , = , = , (3)

      experienced radiologists. Each image carries a PASCAL- VOC XML annotation encoding a lesion box ( min , min , max , max); images are JPEG standardised to 718 × 500 × 3.

      LymphUs [2] comprises cervical lymph node ultrasound from 338 PTC patients, across two independent centres (Centre 1:180 patients, 99 benign / 81 malignant; Centre 2:

      158 patients, 76 benign / 82 malignant), with segmentation masks alongside B-Mode images. Only B-mode images were used; mask, lesion and fusion renderings were excluded,

      2

      Table I

      THROX comparison with previous and recent studies. Reported results are not directly comparable due to differences in datasets, metrics, and test sizes.

      Study

      Yea

      R

      Data

      Endpoint

      TEST

      Dency Flora et al [11]

      2024

      Tabular (yes)

      Dysfunctio n

      –

      Nugroho et al. [18]

      2020

      165 img. (no)

      Margin reg./irreg

      96

      Pimpalakar & Gulame [27]

      2025

      983 lumps (no)

      Benign

      /Malignant

      81

      Rashed et al. [4]

      2025

      Public

      Benign

      /Malignant

      –

      Zhang et al. [7]

      2025

      TN5000(yes

      )

      Detection

      1,00

      0

      AI Pooled [12]

      –

      >23k pts (mixed)

      Nodal Metastasis

      –

      Radiologist s pooled [12]

      –

      >23k pts (mixed)

      Nodal Metastasis

      –

      THYROX

      thyroid (ours)

      2026

      TN5000

      (yes)

      Benign/ Malignant

      750

      THYROX

      nodal (ours)

      2026

      LymphUs (yes)

      Nodal metastasis

      51

      Table II

      Method

      Accuracy

      Nodal

      Survey: hybrid DCNN

      93.85%

      No

      9 Geometric feat

      + MLP

      94.79%

      No

      2 × YOLOv5 +

      NMS + XGBoost

      99.64%

      No

      9 backbones benchmarked

      ResNet50 best

      No

      DETR, Faster R- CNN, SSD

      mAP 0.58

      No

      Meta-analysis, 27 studies

      sens. 80%

      Yes

      Human readers

      Sens. 51%

      Yes

      4-backbone ens

      +TTA

      91.87%

      Yes

      Dataset Characteristics and splits. oversampling was limited to training data, keeping validation and test set unchanged for unbiased evaluation.

      Warm-start DenseNet121 + TTA

      88.24%

      Yes

      Table III

      Augmentation pipeline, model configurations and footprint. model size did not correlate with accuracy: the smallest model performed best, while the largest failed to converge.

      Property

      TN5000[19]

      LymphUs [20]

      Anatomical target

      Thyroid nodule

      Cervical lymph node

      Clinical endpoint

      Benign/malignant

      Non-met. / metastatic

      Images used

      5,000

      338

      Benign / malignant

      1,428 / 3,572

      175 / 163

      Class ratio

      1 : 2.50

      1 : 0.93

      Patients

      Not disclosed

      338

      Centres

      Single release

      2 (180 + 158)

      Modality

      B-mode

      B-mode (+masks, unused)

      Train (ben. / mal.) after balancing

      3,500 (998 /

      2,502)

      233 (120 / 113)

      Validation

      750 (214 / 536)

      54 (28 / 26)

      Test

      750 (214 / 536)

      51 (27 / 24)

      Partition level

      Image

      Patient

      Fig. 1. The THYROX architecture.

      Where compactness equals unity for a circle, convexity < 1 signals an irregular margin, and dispersion quantifies radial irregularity directly.

      Table IV

      Thyroid classification on TN5000 test set ( = 750)

      using three-view TTA.

      Transform

      Thyroid

      Nodal

      Horizontal / Vertical flip

      0.5 / 0.3

      0.5 / 0.5

      Rotation

      ±20°

      ±30°

      Affine translate / scale

      6% / ±10%

      10% / ±15%

      Brightness / contrast jitter

      ±0.20

      ±0.25

      Random erasing

      =0.25

      =0.30

      Mixup [25] rate

      0.5

      0.3

      thyroid, 10 nodal); batch size was 32 and all seeds were fixed at 42.

      A note affecting Fig. 4: on mixup epochs, training accuracy is computed by comparing predictions on the blended batch against the original labels , and is therefore systematically understated. This is why validation accuracy exceeds training throughout, and is a measurement artefact rather than underfitting

      circularity / , solidity A / Ah , elongation b / a, roughness P / Ph, equivalent diameter ration 2(/)/ and margin thickness ( )/, with , the fitted ellipse axes follow analogously.

      1. Architechutres and training

        Four backbones spanning the residual, densely connected, compound scaled and modernized-convolutional families were selected: ResNet50[28], DenseNet121[26], EfficientNetV2-S[16] and ConvNeXt-Tiny[23], all instantiated from ImageNet-pretrained weights via timm with two-way head. Transformer backbones were excluded on the grounds that at 5,000 images the data regime favours convolutional inductive bias. Pre- architecture regularization (Table IV) was set after a preliminary run in which EfficientNetV2-S exhibited the largest train validation gap.

        Because the thyroid classes are imbalanced at 1:2.5, each class c is weighted by inverse frequency, normalized

      2. Cross-anatomical transfer

        The nodal training set contains 233 images, a regime in which ImageNet transfer yields limited benefit [20] and indomain pretraiing repeadtedly helps [12]. Evidence that ultrasound represents generalise across organs [10] motivates our hypothesis: a backbone trained to discriminate benign from malignant thyroid nodules acquires speckle-texture and margin morphology features that transfer to cervical nodal assessment despite the change of organ. Operationally, a DenseNet121 is loaded from the thyroid checkpoint with strict=False, chosen over ResNet50 for parameter economy (27.1 vs 90.0 MB), and trianed with 0 = 5 × 105 , weight decay 2 × 104 and mixup rate 0.3 a gentler schedule that avoids ealry destruction of the trnasferred representation.

      3. Inference, ensembling and risk stratification

        At inference each image is presented under three transforms identity, horizontal flip and vertical flip and the softmax vectors averaged, so that for model

        TTA() = 1 softmax ( ( ())). (9)

        to sum to the number of classes C:

        3

        {I,H,V}

        1/

        =

        (4)

        Architecture level soft voting over the converged backbones then gives the ensemble posterior and decision

        =1

        1 /

        () = 1

        TTA

        (), = [() > ], (10)

        Targets are smoothed by = 0.1 to discourage over-

        =1

        ,mal

        confident logits, = (1 ) + / , giving the

        with = 0.5. the posterior is mapped to an integer risk

        training objective

        score

        100() +

        1

        2

        {0, ,100

        } and thence to three

        =

        log

        ,

        =

        . (5)

        tiers mirroring the structure of TI-RADS [25],

        CE =1

        Low,

        tier() = {Intermediate, <

        High, > .

        (11)

        Mixup [25] is applied with probability perbatch: for a pair of examples (, ), (, ) and Beta(, )

        = + (1 ), (6)

        mix = CE((), ) + (1 )CE((), ), (7)

        With = 0.2 and = 0.5 (thyroid) or 0.3 (nodal). Optimization used AdamW with automatic mixed precision under a cosine-annealed learning rate,

        Boundaries are selected on the validation split only, as the largest score retaining 0.95 and as the smallest reaching 0.90. Tier level results are reserved for future work.

      4. Explainability

        Grad-CAM [17] is computed on the final convolutional block of the selected model. For target class and feature maps of that block, the channel weights and the localization map are

        0 max

        = 1 (1 + cos(/ )), (8)

        2

        =

        1

        , = ReLU ( ), (12)

        Where = 25 (thyroid) or 30 (nodal). Model selection used best validation accuracy with early stopping (patience 7

        Where is the number of spatial positions. Because TN5000 supplies a ground truth box for every image,

        attention quality is quantifies rather than merely displayed. Over test images we report the pointing game [14] and the energy-in-box fraction,

    2. Two negative results

      ConvNeXt-Tiny did not converge. Its training remained At 0.49-0.51 for all eight epochs before early stopping,

      PG =

      1 [arg max (, ) ], (13)

      validation accuracy oscillated between the class priors

      =1

      (,)

      (0.7147 and 0.2853), and it assigned the malignant label to all

      750 test images yielding accuracy exactly equal to majority class prevalence and ROC-AUC 0.5601 (Fig. 4). We attribute

      1 (, )

      (,)

      =

      =1

      (, ), (14)

      this to optimizer architecture mismatch rather than a deficiency of the architecture mismatch rather than a deficiency of the architecture: the shared protocol applies

      Together with the IoU of (2) between and thresholder

      at half its maximum.

      Table V

      TN5000 Confusion matrices and error costs. Ensemble achieved 91.87% Accuracy, 95.34% Sensitivity and 83.18 Specificity, Slightly outperforming DENSENET121

      AdamW at 0 = 3 × 104 with uniform weight decay, whereas ConNeXt-family models are conventionally trained with far lower peak rates, layer-wise decay, stochastic depth and much longer schedules [15]. At 5,000 images and 25 epochs none of those conditions held. Identical regularization protocols therefore do not transfer across architecture families directly relevant to any multi-backbone benchmark under a shared recipe.

      Model

      TN

      FP

      FN

      TP

      Missed

      Unnec.

      ResNet50

      178

      36

      34

      502

      6.3%

      16.8%

      DenseNet121

      177

      37

      25

      511

      4.7%

      17.3%

      EfficientNetV2- S

      175

      39

      32

      504

      6.0%

      18.2%

      ConvNeXt-

      Tiny

      0

      214

      0

      536

      0.0%

      100%

      Ensemble

      178

      36

      25

      511

      4.7%

      16.8%

      Single-box regression proved inadequate for localization. Validation Smooth L1 loss fell monotonically from 0.0078 to

      0.0049, so optimization succeeded, but localization quality

      did not: mean IoU 0.2859, with only 23.73% of predictions exceeding IoU 0.5. The failure is structural: regressing four coordinates form a global embedding presumes a single lesion of predictable scale and position, with no mechanism for scale invariance, competing candidate regios, or confidence ranked suppression. Detectors on this dataset reach mAP@0.5 0.58 [7] and dedicated single stage detectors elsewhere exceed 0.70 [27]. The consequence is visible in Fig. 7, where the predicted box sits well within the true lesion extent. Because region-of-interest extraction, geometric features and the hybrid classifier all depend on a reliable lesion region, this stage is keystone of the pipeline,

      1. Evaluation metrics

      With malignant as the positive class and (TP, FP, FN, TN) the confusion-matrix counts, we report accuracy, precision TP/(TP+FP), specificity TN/(TN+FP), negative predictive value TN/(TN+FN), F1 = 2TP/(2TP+FP+FN),

      Youdens J = Sens +Spec -1, and balanced accuracy 1(Sens

      2

      + Spec). ROC-AUC and PR-AUC are computed from the

      ensemble posteriors of (10). Because the nodal test set contains only 51 images, interval estimates use the Wilson score method rather than the normal approximation: for successes in trials with = / and = 1.96,

      not a peripheral component.

    3. Cervical nodal metastasis

      The warm-started DenseNet121 reached best validation

      accuracy 0.8704 at epoch 8, with early stopping at epoch 17. Test performance appears in Table VI and Fig. 5. The model made no false-positive predictions: every benign node was correctly identified, giving 100% specificity and precision, at the cost of six missed metastases of 24(75.00% sensitivity). At 51 test images the intervals are necessarily wide sensitivity spans 55.10-88.00% and the point estimates should be read

      2

      (1 ) 2

      accordingly rather than as precise performance claims.

      CI95% =

      + 2 ±

      + 42

      2

      . (15)

      1 +

      Localization is evaluated by mean IoU and IoU@0.5.

      Experiments ran on a single NVIDIA Tesla T4 (16GB) with PyTorch 2.11 and CUDA 12.8.

  4. Results

    A. Thyroid nodule classifications

    The performance of the four backbone models and the ensemble on the TN5000 test set with three view test-time augmentation (TTA) is evaluated using confusion matrices and Wilson confidence intervals. The corresponding results are visualised in the Figs. 2,3 and 6

    Fig. 2. Confusion matrices for the TN5000 test set.

    Fig. 3. Receiver operating characteristic (left) and precision recall (right) curves on the TN5000 test set

    Table VI

    LYMPHUS TEST SET CLASSIFICATION (n=51) USING WARM STARTED DENSENET121 WITH TTA. WILSON 95% CONFIDENCE INTERVALS

    Metric

    Value

    95% CI

    Accuracy

    88.24%

    76.62-94.50

    Precision

    100.00%

    –

    Sensitivity

    75.00%

    55.10-88.00

    Specificity

    100.00%

    87.54-100.00

    Negative

    Predictive Value

    81.82%

    –

    F1-Score

    85.71%

    –

    Balanced

    accuracy

    87.50%

    –

    Confusion

    Matrix

    TN 27, FP

    FN 6, TP 18

    1. Explainability and end-to-end output

      Fig. 7 shows Grad-CAM and the complete pipeline output for an unseen transverse acquisition. Attention concentrates on the isthmic nodule, which the model classifies as malignant with 90.8% confidence the desired behaviour. The same map, however, contains a second high-activation region at the inferior border, over an acoustic-shadow area posterior to the lesion that corresponds to no nodular structure. We report this explicitly: it shows that Grad-CAM attention here is not fully lesion-specific, consistent with benchmarking evidence from other modalities [14], and it is the direct motivation for quantifying attention

    2. Comparision with published methods

    Table I places these results beside published work. Against pooled radiologist sensitivity of 51% for cervical nodal metastasis across 27 studies [6], the nodal modules 75% from 233 training images is a meaning margin, though the 51 images test set precludes a strong claim. Against the pooled AI sensitivity of 80% and specificity of 83% in the same meta- analysis, THYROX sits below on sensitivity and above on specificity, reflecting the threshold asymmetry noted above.

  5. Discussion

    1. Why the smallest model won, and what that implies

      DenseNet121 achieved the highest accuracy, sensitivity and F1 with the smallest checkpoint by a factor of three to

      four. A plausible mechanism is architectural: dense connectivity keeps early-layer representations directly available at the head, and in ultrasound the discriminative features margin irregularity, microcalcification, subtle echogenicity gradients are fine-grained and high-frequency, precisely the information deep residual bottlenecks discard. Feature reuse rather than capacity appears to be the useful property, which also explains why the largest model performed worst. The broader implication is that at few thousand images the training recipe dominates architectural choice: three architectures with different inductive biases converged to within 0.002 AUC under one protocol, with the same protocol applied to a fourth failed completely.

      This bounds what our results support. The 0.0021 ROC- AUC range across three backbones on 750 test images is far smaller than the sampling variability such a text set admits, and we performed neither DeLong testing nor repeated seed runs; we therefore do not claim any of three is superior. The caution applies with greater force to the nodal results, where a sensitivity interval spanning 55-88% is compatible both with clinically useful performance and with performance that would not be.

    2. The operating point matters more than the accuracy

      The clinacally weakest number here is thyroid specificity at 83.18% approximately one benign nodule in six is lfagged malignant, which in an FNA triage aid translates into un- necessary biopsies the burden that motivated weighted at

      2.5:1 while screening populations are overwhelmingly benign, postive predicitve value would fall substantially in deployment. The 0.5 treshold is arbirary: tuning tier boundaries on validation data converts it into a deliberate operating point, and a tiered output that defers explicitly is more clinically honest than a binary label.

      The nodal module shows the opposite asymmetry and need the opposite correction. Perfect specifity with 75% sensitivity eans it never over calls metastasis but misses one in four. For nodal staging that is the wrong trade-off a missed metastssis may lead to indequate initial suregery, whereas a false positve prompts further imaging so the theshold should move toward sensitivity.

      Fig. 4. Training, validation loss and accuracy

      Fig. 5. Confusion matrix for lymph node classification

      Fig. 6. Comparative performance. Top: accuracy, sensitivity, specificity and F1 by model. Bottom: ROC-AUC and PR-AUC by model

    3. Positioning, limitations and future work

    THYROX is a triage and second-reader aid, a framing with a stronger base than autonomus operation [19], [9]: the tiered output of (11), a visual explaination permitting a reader

    Fig. 7. Qualitative output. Top: original image, Grad-CAM heat map and overlay with predicted class. Bottom: end-to-end output.

    Fig. 8. Risk stratification. (a) Probability-to-risk mapping. (b) Risk score in the THYROX interface.

    to reject an implausible prediction, and a fooprint compatible with modest harware all fit that role.

    The limitations are substantial. All results derive from retrospective secondary analysis of public data; thyroid partitioning is image-level because the release does not expose patient identifiers; and the thyroid model has no external validation. The nodal test set contains 51 images, giving intervals too wise for precise claims. Only B-mode imaging was used elastography, Doppler and the LymphUs masks were not exploited. Results come from a single partition and seed without cross-validation or statistical testing, and the partition used (3,500/750/750) differs from TN5000s official 3,500/500/1,000, limiting comparability. The endpoint is binary, and no comparison against radiologists on these images was performed.

    Immediate priorities are therefore an ablation isolating ImageNet from thyroid-ultrasound initialization, repeated seed runs with DeLong testing, and re-evaluation on the official partition. Beyond these, LymphUss two-centre structure permits genuine externally validation by training on Centre 1 and testing on Centre 2 the strongest credibility improvement available at low cost while the distributed masks would support a segmentation stage far more reliable than box regression, enabling the geometric feature and hybrid classifier stages. Multi-view fusion [27], multimodal fusion [5] and ultrasound foundation model backbones [10]

    are natural extensions; a prospective reader study would be required before any deployment claim.

  6. CONCLUSION

We presented THYROX, a unified explainable system addressing both components of PTC ultrasound assessment nodule malignancy and cervical nodal status in one pipeline using exclusively public, biopsy confirmed data. A soft-vote ensemble of three convolutional backbones with test time augmentation achieved 91.87% accuracy, 95.34% sensitivity, 83.18% specificity and ROC-AUC 0.9675 on the TN5000 test set, and a nodal classifier warm started from the thyroid backbone reached 88.24% accuracy from only 233 training images indicating that in-domain ultrasound representations offer a practical route to usable performance under label scarcity. Backbone choice within the convolutional family proved immaterial and the smallest model performed best, wile ConvNeXt-Tiny under the shared recipe and single-box regression for localization both failed in ways we report rather than omit. THYROX is positioned as a triage and second reader aid; its specify, absence of external validation and small nodal evaluation preclude any stronger claim

This study is a secondary analysis of publicly released, de-identified datasets; no new patients data were collected and institutional review board approval was not required

REFERENCES

  1. D. W. Chen and M. R. Haymart, Unravelling the rise in thyroid cancer incidence and adressing overfiagnosis, Nat. Rev. Endocrinol., vol. 22, no.1, pp. 10-20, 2026.

  2. A. Mohammadi et al., lyumphUs: a mulitcenter open-access database of lymph node ultrasound images in patients with papillary thyroid carcinoma, Data Brief, vol. 66, art 112694, 2026.

  3. G. N. Sujini and S. Balakrishna, automated thyroid nodule classification in ultrasound imaging using a hybrid vision transformer and Wasserstein GAN with gradient penalty, Sci. Rep., vol 15, art. 40786, 2025.

  4. A. Rashed, T. Medhat, and A. Elgarayhi enhancing automatic diagnosis of thyroid nodules from ultrasound scans leveraging deep learning models, Sci. Rep., vol. 15, art. 4036, 2025.

  5. P. Shen et al., Explainable multimodal deep learning for predicting thyroid cancer lateral lymph node metastasis using ultrasound imaging, Nat. Commun., vol. 16, art. 7052, 2025.

  6. X. Wang et al., Ultrasound based artificial intellignece for predicting cervical lymph node metastsis in papillary thyroid cancer: a systematic reviw and meta-analysis, Front. Endocrinol., vol. 16, art. 1570811, 2025.

  7. H. Zhang, Q Liu, X. Han, L. Niu, and W.Sun, TN5000: and ultrasound image dataset for thyroid nodule detection and classification, Sci. Data, vol.12, art. 1437, 2025.

  8. A. Pimpalkar and M. Gulame, Thyroid lumps detection in ultrasound images utilizing muli-view deep learning approach, in Prox. IEEE Int. Conf. Recent Adv. Ciout. Syst. (REACS), 2025.

  9. D. Xu et al., The clinical value of artifical intelligence in assisting junior radiologists in thyroid ultrasound: a multicenter prospective study, BMC Med., vol. 22, art. 293, 2024.

  10. J. Jiao et al., USFM: a universal ultrasound foundation model generalized to tasks and organs towards label efficient image analysis, Med. Image Anal., vol. 96, art. 103202, 2024.

  11. G. Dency Flora et al., Identification of thyroid disorder using deep learning techniques a survey, in Proc. 10th Int. Conf. Adv. Comput. Commun. Syst.(ICACCS), 2024.

  12. S. Azizi et al., Robust and data-efficient generalization of self- supervised machine learning for diagnostic imaging, Nat. Biomed. Eng., vol. 7, no. 6, pp. 756-779, 2023.

  13. W. Li et al., Systematic review and meta-analysis of ACR TI-RADS inter-reader reliabilty for risk stratification of thyroid nodules, Front. Oncol., vol. 12, art.840516, 2022.

  14. A. Saporta et al., Benchmaking saliency methods for chest X-ray interpretation, Nat. Mach.Intell., vol. 4, no. 10, pp. 87=67-878, 2022.

  15. Z. Liu et al., A ConvNet for the 2020s, in Proc. IEEE/CVF CVPR, 2022, pp. 11966-11976.

  16. M. Tan and Q. V. Le, EffiicientNetV2: smaller models and faster training, in Proc. ICML, PMLR vol. 139,2021, pp. 10096-10106.

  17. R. R Selvaraju et al., Grad-CAM: visual explanations from deep networks via gradient-based localization, Int. J. Comput. Vis., vol. 128, no. 2, pp. 336-359, 2020.

  18. H. A Nugroho, E. l. Frannita, and A. H. T. Hutami, Thyroid nodules catego rization based oargin features using deep learning, in Proc. 3rd Int. Semin. Res. Inf. Technol. Intell. Syst. (ISRITI), 2020, pp. 499-504.

  19. M. Budha et al., Management of thyroid nodules seen on US images deeo kearing may match performance of radiologists, Radiology, vol. 292, no. 3, pp. 695-701, 2019.

  20. M. Raghy, C. Zhang, J. Kleinberg, and S. Bengio, Transfusion: understanding tranfer learning for medical imaging, in Adv. Neural Inf. Process. Syst., vol. 32, 2019.

  21. J. K. Hoang et al., Reduction in thyroid nodule biopsies and improved accuracy with ACR TI-RADS, Radiology, vol.287, no. 1, pp. 185- 193, 2018.

  22. J. H. Lee et al., Deep learning-based computer-aided diagnosis system for localization and diagnosis of metastatic lymph nodes on ultrasound: a pilot study, Thyroid vol. 28, no. 10 pp. 1332-1338, 2018.

  23. H. Zhang, M. Cisse, Y. N. Dauplin, and D. Lopez-Paz, mixup: beyond epirical risk minimization, in Proc. ICLR, 2018.

  24. E. S. Cibas and S. Z. Ali, The 2017 Bethesda System for Reporting Thyroid Cytopathology, Thyroid, vol. 27, no. 11, pp. 1341-1346,

    2017.

  25. F. N. Tessler el al.,ACR Thyroid Imaging, Reporting and Data System (TIRADS): white paper of the ACR TI-RADS Committee, J. Am. Coll. Radiol., vol. 14, no. 5, pp. 587-595, 2017.

  26. G. Huang, Z. Liu, L. van der Maaten, and K. Q. Qeinberger, Densely connected convolutional networks, in Proc. IEEE CVPR, 2017, p. 2261-2269.

  27. B. R. Haufen et al., 2015 American Thyroid Association management guidelies for adult patients with thyroid nodules and diffrentiated thyroid cancer, Thyroid, vol. 26, no. 1, pp. 1-133, 2016.

  28. K. He, X. Zhang, S. Ren, and J. Sun, Deep residual learning for image recognition, in Proc. IEEE CVPR, 2016, pp.770-778.