Pith. sign in

REVIEW 4 major objections 6 minor 51 references

ThyroidEffi 1.0: A Cost-Effective System for High-Performance Multi-Class Thyroid Carcinoma Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read The paper claims a three-part pipeline — YOLOv10-guided augmentation, curriculum training from crops to full images, and EfficientNetB0 — classifies thyroid FNAB images at macro F1 89.19%, with external AUCs of 0.95, 0.74, and 0.84.

desk verdict A cost-conscious and honestly deployed thyroid FNAB system with real external validation, but the internal headline metric is suspect because architecture selection used the test set, and the external numbers are weaker than the abstract admits. read the letter →

arxiv 2504.14139 v2 pith:HXGE3JWX submitted 2025-04-19 cs.CV cs.AI

classification cs.CVcs.AI
keywords thyroidcarcinomafine-needleaspirationbiopsymedicalimageclassificationmulti-classBethesdasystemcost-effectiveEfficientNetB0curriculumlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that accurate multi-class classification of thyroid fine-needle aspiration biopsy (FNAB) images does not require a heavy deep-learning stack. Its pipeline combines three ideas — detector-guided augmentation with YOLOv10, a curriculum training order that moves from local crops to full images, and the 4-million-parameter EfficientNetB0 backbone — and targets the three categories that actually steer treatment: Benign (Bethesda II), Indeterminate/Suspicious (Bethesda I, III, IV, V), and Malignant (Bethesda VI). On the internal test set the Basic model reaches macro F1 of 89.19%, the average of the three classes' F1 scores, with per-class AUCs of 0.98, 0.95, and 0.96, and on 1,015 independent images from a second hospital it reports AUCs of 0.9495, 0.7436, and 0.8396. If true, the practical payoff is a triage tool that runs 1,000 cases in about 30 seconds on a 12-core CPU.

What carries the argument

The central machinery is the four-module pipeline M1–M4, whose job is to give one 1024×768 FNAB field both local and global scale information while keeping inference cheap. M1 is the augmentation engine: a YOLOv10 detector (trained on 120 annotated training images, used only during training) locates cell clusters of at least ten cells and drives five augmentation sets — the original image (A), the original overlaid with the top-eight detected clusters (B), eight dense-cluster crops (C), and twenty-four 256×256 grid crops split into sets D and E — expanding the training data by a factor of 34. M2 is the curriculum: within every epoch the augmented views are fed in the order E, D, C, B, A, so the model first learns from small noisy patches and ends on full images, exploiting the recency bias of gradient updates. M3 is the model choice: EfficientNetB0 (about 4 million parameters), selected after comparing twelve models from six families on macro F1, parameter count, and inference time. M4 is the multi-region head: a Transformer-inspired module that partitions the image into a 3×4 grid, treats each patch as a token, and initializes the classification token with the backbone's full-image output, so local details are contextualized globally in a single pass.

What would settle it

Have a cytopathologist reread the external cases the model labelled Malignant but whose reference diagnosis is Indeterminate (128 of 315): if most of those images contain no truly malignant cells, the crop-label-inheritance strategy is injecting label noise and the reported scores partly measure that noise. A quantitative check is to retrain the Basic model with the curriculum order reversed while keeping everything else fixed — if internal-test macro F1 stays at 89.19%, the E→D→C→B→A ordering is not carrying the gain.

Watch

Extended reading notes

Core claim

Stated the way a sympathetic reader would state it: a deliberately cost-efficient training recipe can produce a clinically usable three-way thyroid FNAB classifier. The recipe is (1) train a YOLOv10 detector on 120 annotated training images to find dense cell clusters, then use its bounding boxes to build a ×34 augmentation — original image, an overlay of the top-eight clusters, eight high-density crops, and 24 grid crops; (2) present those views inside each epoch in the order E→D→C→B→A, from noisy local patches to the full image; (3) use EfficientNetB0 as the backbone; and (4) for the Premium variant, add a Transformer-inspired head that treats a 3×4 grid of patches as tokens and initializes the class token with the full-image encoding. The reported outcome is macro F1 of 89.19% (Basic) and 89.77% (Premium) on the internal test set, per-class AUCs of 0.98/0.95/0.96, and external AUCs of 0.9495/0.7436/0.8396 on 1,015 images collected prospectively at a second site. The authors' conclusion is that high-accuracy, interpretable thyroid FNAB classification is achievable with minimal computational demands.

Load-bearing premise

The load-bearing premise is that one 1024×768 FNAB field per patient, captured at 40× with Diff-Quick staining, is representative enough to label the patient's cytology, and that every cropped sub-region generated from that image can inherit the full-image diagnostic label as training truth.

Editorial extensions

If this is right

  • A triage-grade FNAB classifier can run on hardware already present in modest clinics: about 1,000 cases in 30 seconds on a 12-core CPU, with only the raw image as input at inference and no manual region extraction or ensemble.
  • In the paper's own comparison, the augmentation and curriculum protocol raises macro F1 for eleven of the twelve backbones tested (EfficientNetB0 from 0.8555 to 0.8919), so most of the gain sits in the training recipe rather than in any single architecture.
  • The three-way clinical mapping — Benign (Bethesda II), Indeterminate/Suspicious (Bethesda I, III, IV, V), Malignant (Bethesda VI) — aligns the classifier's output directly with the treatment decision, which is what lets the system act as a second reader rather than a research tool.
  • The accuracy claim is conditional on class and site: the external AUC holds for Benign (0.9495) and Malignant (0.8396) but drops to 0.7436 for the Indeterminate/Suspicious group, with external macro F1 at 0.68, so the deployable promise is strongest for the confident classes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the paper leaves implicit: since the external errors concentrate on Indeterminate cases being called Malignant, feeding several FNAB fields per patient and aggregating their probabilities by average would likely lift the 0.7436 external AUC without any retraining.
  • Because every crop inherits the full-image label, sets C, D, and E carry deliberately noisy truth; if per-crop cytopathologist labels existed, the same pipeline would probably gain more from them than from any further architecture change — a comparison the paper does not run.
  • The curriculum's reliance on recency effects in gradient updates implies the E→D→C→B→A order should matter more on small datasets; the same recipe therefore ought to transfer to other scarce-data cytology tasks such as breast FNA, which the paper's own roadmap anticipates.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper presents ThyroidEffi, a three-class (Benign, Indeterminate/Suspicious, Malignant) deep-learning pipeline for classifying thyroid FNAB images. The pipeline combines YOLOv10-based cell-cluster detection for targeted data augmentation, a curriculum-learning training order (crop sets E, D, C, B, A), a lightweight EfficientNetB0 backbone for the Basic model, and a Transformer-inspired multi-region module for the Premium model. The authors report an internal test-set macro F1 of 89.19% with AUCs of 0.98/0.95/0.96, and an external validation on 1,015 images from a different hospital with AUCs of 0.9495/0.7436/0.8396 and a macro F1 of 0.68. The paper claims that high-accuracy, interpretable, cost-effective thyroid FNAB classification is achievable, and it provides code and deployment details.

Significance. If the results were unbiased, the paper would provide a useful, low-cost, deployable system for a clinically meaningful three-way triage task, with the valuable feature of an independent external validation set. The main contributions are the YOLOv10-based augmentation strategy, the curriculum ordering, the lightweight architecture choice, and the prospective external evaluation, all of which are concrete and reproducible. The availability of source code and the explicit reporting of external F1/AUC are strengths. However, the central internal-performance claim appears to be affected by test-set-based architecture selection, and the external results are substantially weaker than the abstract's 'high-accuracy' characterization, particularly for the Indeterminate/Suspicious class. The paper's significance therefore hinges on whether the authors can provide an unbiased estimate of internal performance and reframe their claims to match the external evidence.

major comments (4)
  1. [§3.1, Table 2] The architecture selection appears to have used the held-out test set. Table 2 is explicitly titled 'F1 Score on Test Set for Different Models and Configurations', and EfficientNetB0 is 'designated as ThyroidEffi Basic' after achieving the highest augmented F1 (0.8919) in that same table. Since no validation-set architecture comparison is reported, the reported 0.8919 macro F1 is the maximum over 12 test-set draws and is expected to overstate the true performance of the chosen model. The authors should redo model selection using only the validation set (as described in Appendix E.3 for hyperparameters) and report the test performance of the validation-selected model, or at minimum clearly disclose this selection bias and temper the internal-performance claims accordingly.
  2. [Abstract; §3.3, Table 4] The external validation results do not support the abstract's 'high-accuracy' phrasing for all classes. Table 4 reports external macro F1 of 0.68 and an Indeterminate/Suspicious F1 of only 0.49, and Table 5 shows that 128 of 315 Indeterminate/Suspicious cases were misclassified as Malignant. The abstract and conclusions should be revised to state class-specific limitations and to characterize the system's performance as promising but with meaningful degradation on the indeterminate/suspicious category, rather than as uniformly high accuracy.
  3. [§2.1, Table 1, Appendix C.4] There is a numerical inconsistency in the reported class distribution. Section 2.1 states '871 (Malignant)' and a total of 1,804 images, but Table 1 sums to 781 Malignant (541+125+115) and a total of 1,804, and Appendix C.4 also lists 781 Malignant. Because the class balance is central to interpreting the weighted loss and the reported metrics, this discrepancy must be corrected and verified.
  4. [§2.2.2, §2.1] The crop-label inheritance assumption is load-bearing for the training procedure. Each of the 1,804 images corresponds to a unique patient and a single 1024x768 field, and every crop in sets C, D, and E is assigned the full-image diagnostic label. If a crop contains only benign material from a malignant case, or if the single chosen field misses the diagnostic region, the training labels are noisy. The paper should either provide an analysis of crop-level label quality (e.g., agreement of model predictions on crops vs. full images) or explicitly frame this as a limitation that may weaken the clinical claim.
minor comments (6)
  1. [§3.1] The sentence 'Detailed comparison results ... in the analysis presented in Section 3.1result on test set' contains a typo ('3.1result') that should be corrected.
  2. [Figure 2] The caption refers to 'the first [6] and second [29]' but the surrounding text mentions two research groups; please clarify whether [6] and [29] are both from the same U.S. group and label the figure entries consistently.
  3. [§5.1] The heading 'Feature research' appears to be a typo for 'Future research'; please correct it.
  4. [§3.4, Table 5] The text reports '1000 cases in 30 seconds' while the external set contains 1,015 images; please clarify whether 'cases' here means images or patients, and state the hardware configuration more precisely.
  5. [Appendix E.3] The hyperparameter search is described as conducted on the validation set, but the architecture selection in §3.1 is reported on the test set; aligning the two selection protocols would remove ambiguity about which set informed model choices.
  6. [Results] No confidence intervals or uncertainty estimates are reported for the F1 scores or AUCs; reporting these would make the magnitude of the differences between models (e.g., 0.8919 vs. 0.8690) easier to assess.

Circularity Check

1 steps flagged · score 6.0 of 10

Internal macro F1 is the test-set maximum used to choose ThyroidEffi Basic, making the headline internal result a selected-order statistic rather than an independent prediction.

  1. fitted input called prediction [Section 3.1, Table 2; Section 2.2.3 (M3)]
    "Table 2: F1 Score on Test Set for Different Models and Configurations ... EfficientNetB0, trained with the augmented dataset, achieved the highest macro F1 score of 89.19% and is designated as ThyroidEffi Basic."

    The internal headline metric is the same test-set statistic that selected the model. EffNetB0 is designated as ThyroidEffi Basic because it achieved the highest augmented macro F1 in Table 2, which is explicitly a test-set comparison; the paper then reports that same 0.8919 as the achieved internal result. Therefore 0.8919 is, by construction, the maximum of twelve test-set F1 draws, not an unbiased evaluation of a pre-specified model. The external set was not used for selection, so it retains independent evidentiary value; but the abstract's 'macro F1 of 89.19%' reduces to the model-selection criterion rather than to an independent prediction.

full rationale

The paper is primarily an empirical pipeline paper, and most of its components are standard or externally evaluated: YOLOv10-based augmentation, curriculum-style training order, EfficientNetB0, and a Transformer-inspired module are not justified by self-citation, and no definitional equivalence or self-citation loop was found. The one substantive circularity is the internal test-set evaluation. Section 2.2.3 says twelve models were assessed using criteria including F1-score, and Section 3.1/Table 2 reports those F1 scores on the test set; EffNetB0 was then designated as ThyroidEffi Basic because its augmented test F1 (0.8919) was the highest. Reporting the same 0.8919 as the achieved macro F1 makes the headline internal number a selected maximum over twelve test-set evaluations, which is optimistically biased by construction. Appendix E.3 describes validation-based hyperparameter selection, but the architecture/model-configuration comparison that determines the reported internal result is explicitly on the test set. The external validation set was not used for selection and therefore provides independent, albeit weaker, evidence; this is why the circularity is partial rather than total. The crop-label inheritance concern is a data-quality and label-noise risk, not a circularity, because test and external evaluations are performed on full images.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The system rests on standard ML components and a set of clinical and data-labeling assumptions. The main free parameters are training hyperparameters selected by validation, plus class weights and the augmentation factor. No new physical or conceptual entities are introduced. The most fragile assumptions are the single-image-per-patient representativeness and the assignment of full-image labels to crops.

free parameters (7)
  • Learning rate = 1e-4
    Selected via grid search over [1e-5, 1e-3] on validation macro F1 (Appendix E.3).
  • Batch size = 120
    Grid search over [12, 24, 36, 120]; best validation macro F1 (Appendix E.3).
  • Weight decay = 1e-3
    Grid search over [1e-4, 1e-2]; selected on validation (Appendix E.3).
  • Dropout rate = 0.2
    Grid search over [0.1, 0.5]; selected on validation (Appendix E.3).
  • Transformer encoder layers (Premium) = 5
    Grid search over [2, 5]; selected on validation (Appendix E.3).
  • Class weight w_j = total/(num_classes x freq(j))
    Computed from training class frequencies to reweight cross-entropy (Appendix E.2); depends on the training distribution.
  • Augmentation factor = 34
    Design choice in M1: 1 (A) + 1 (B) + 8 (C) + 12 (D) + 12 (E) crops per image; chosen by the authors, not independently justified.
assumptions (4)
  • domain assumption The TBSRTC-consensus labels assigned by cytopathologists are accurate enough to serve as ground truth.
    Section 2.1 and Appendix B.2 describe consensus review, while Section 4.4 acknowledges inter-observer variability, so the labels are not a definitive gold standard.
  • domain assumption Grouping Bethesda II as Benign, I/III/IV/V as Indeterminate/Suspicious, and VI as Malignant matches clinical treatment decisions in Vietnam.
    Section 1.4 defines the three groups; if this mapping does not reflect real management, the classification targets are mis-specified.
  • domain assumption A single 1024x768 FNAB image per patient is representative of that patient's cytology.
    Section 2.1 states 'each corresponding to a unique patient'; the model sees only one image per patient, so unrepresentative images would break the clinical claim.
  • ad hoc to paper Every augmented crop produced from an image inherits the full-image diagnostic label.
    Section 2.2.2 explicitly assigns the full image's label to localized crops; crops without diagnostic material introduce label noise, and this assumption is specific to the proposed augmentation scheme.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ThyroidEffi 1.0: A Cost-Effective System for High-Performance Multi-Class Thyroid Carcinoma Classification." pith.science (2026). https://pith.science/paper/HXGE3JWX

@misc{pith2026250414139,
  author       = {Pith},
  title        = {Pith review of: ThyroidEffi 1.0: A Cost-Effective System for High-Performance Multi-Class Thyroid Carcinoma Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HXGE3JWX}},
  note         = {Machine review of arXiv:2504.14139}
}
read the original abstract

Background: Automated classification of thyroid Fine Needle Aspiration Biopsy (FNAB) images faces challenges in limited data, inter-observer variability, and computational cost. Efficient, interpretable models are crucial for clinical support. Objective: To develop and externally validate a deep learning system for multi-class thyroid FNAB image classification into three key categories directly guiding post-biopsy treatment in Vietnam: Benign (Bethesda II), Indeterminate/Suspicious (BI, III, IV, V), and Malignant (BVI), achieving high diagnostic accuracy with low computational overhead. Methods: Our pipeline features: (1) YOLOv10 cell cluster detection for informative sub-region extraction/noise reduction; (2) curriculum learning sequencing localized crops to full images for multi-scale capture; (3) adaptive lightweight EfficientNetB0 (4M parameters) balancing performance/efficiency; and (4) a Transformer-inspired module for multi-scale/multi-region analysis. External validation used 1,015 independent FNAB images. Results: ThyroidEffi Basic achieved macro F1 of 89.19% and AUCs of 0.98 (Benign), 0.95 (Indeterminate/Suspicious), 0.96 (Malignant) on the internal test set. External validation yielded AUCs of 0.9495 (Benign), 0.7436 (Indeterminate/Suspicious), 0.8396 (Malignant). ThyroidEffi Premium improved macro F1 to 89.77%. Grad-CAM highlighted key diagnostic regions, confirming interpretability. The system processed 1000 cases in 30 seconds, demonstrating feasibility on widely accessible hardware. Conclusions: This work demonstrates that high-accuracy, interpretable thyroid FNAB image classification is achievable with minimal computational demands.

Figures

Figures reproduced from arXiv: 2504.14139 by the authors.

Figure 1
Figure 1. Graphical Abstract 1 Introduction 1.1 Background and Significance Thyroid carcinoma is a prevalent endocrine malignancy worldwide [1]. Accurate and efficient diagnosis of thyroid conditions is crucial for effective treatment and improved patient outcomes. FNAB is a minimally invasive procedure widely used for assessing thyroid abnormalities and guiding treatment decisions [2]. Historically, FNAB diagnosis has relied… view at source ↗
Figure 2
Figure 2. Test results from the limited number of studies classifying more than two classes are presented. From left [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Workflow of thyroid nodule diagnosis, from clinical examination and ultrasound-guided biopsy to slide image [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visual overview of our methodology, encompassing data pre-processing, model training with novel techniques [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of data distribution before and after applying the ThyroidEffi Basic model for classifying Benign, [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Grad-CAM of a slide image (left) and 12 sub-images (right) after pass ThyroidEffi Basic [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 41 canonical work pages

  1. [1]

    Resistance of papillary thyroid cancer stem cells to chemotherapy

    Raffaella Giuffrida, Luana Adamo, Gioacchin Iannolo, Luisa Vicari, Dario Giuffrida, Adriana Eramo, Massimo Gulisano, Lorenzo Memeo, and Concetta Conticello. Resistance of papillary thyroid cancer stem cells to chemotherapy. Oncology letters, 12(1):687–691, 2016

  2. [2]

    Diagnostic accuracy of fine needle aspiration biopsy versus postoperative histopathology for diagnosing thyroid malignancy

    Aziz Aliyev, Irada Aliyeva, Francesco Giammarile, Narmin Talibova, Gunay Aliyeva, and Fuad Novruzov. Diagnostic accuracy of fine needle aspiration biopsy versus postoperative histopathology for diagnosing thyroid malignancy. Endocrinology, Diabetes & Metabolism, 5(6):e373, 2022

  3. [3]

    Overview of diagnostic error in health care

    Erin P Balogh, Bryan T Miller, John R Ball, Engineering National Academies of Sciences, Medicine, et al. Overview of diagnostic error in health care. In Improving diagnosis in health care. National Academies Press (US), 2015. 15 A PREPRINT

  4. [4]

    Medical image analysis using deep learning algorithms

    Mengfang Li, Yuanyuan Jiang, Yanzhou Zhang, and Haisheng Zhu. Medical image analysis using deep learning algorithms. Frontiers in Public Health, 11:1273253, 2023

  5. [5]

    Artificial intelligence in cytopathology: A neural network to identify papillary carcinoma on thyroid fine-needle aspiration cytology smears

    Parikshit Sanyal, Tanushri Mukherjee, Sanghita Barui, Avinash Das, and Prabaha Gangopadhyay. Artificial intelligence in cytopathology: A neural network to identify papillary carcinoma on thyroid fine-needle aspiration cytology smears. Journal of Pathology Informatics, 9(1):43, 2018. ISSN 2153-3539

  6. [6]

    Thy- roid cancer malignancy prediction from whole slide cytopathology images

    David Dov, Shahar Z Kovalsky, Jonathan Cohen, Danielle Elliott Range, Ricardo Henao, and Lawrence Carin. Thy- roid cancer malignancy prediction from whole slide cytopathology images. In Machine Learning for Healthcare Conference, pages 553–570. PMLR, 2019

  7. [7]

    Elliott Range, David Dov, Shahar Z

    Danielle D. Elliott Range, David Dov, Shahar Z. Kovalsky, Ricardo Henao, Lawrence Carin, and Jonathan Cohen. Application of a machine learning algorithm to predict malignancy in thyroid cytopathology. Cancer Cytopathology, 128(4):287–295, 2020

  8. [8]

    Computer-assisted fine-needle aspiration cytology of thyroid using two-stage refined convolutional neural network

    Wensi Duan, Lili Gao, Juan Liu, Cheng Li, Peng Jiang, Lang Wang, Hua Chen, Xiaorong Sun, Dehua Cao, Baochuan Pang, Rong Li, and Sai Liu. Computer-assisted fine-needle aspiration cytology of thyroid using two-stage refined convolutional neural network. Electronics, 11(24), 2022. ISSN 2079-9292

Show all 51 references
  1. [9]

    Lee, Russel Kahmke, Daniel Rocke, Jonathan Cohen, Ricardo Henao, Lawrence Carin, and Danielle Elliott Range

    Serge Assaad, David Dov, Richard Davis, Shahar Kovalsky, Walter T. Lee, Russel Kahmke, Daniel Rocke, Jonathan Cohen, Ricardo Henao, Lawrence Carin, and Danielle Elliott Range. Thyroid cytopathology cancer diagnosis from smartphone images using machine learning. Modern Patholog...

  2. [10]

    Rocke, Russel R

    David Dov, Danielle Elliott Range, Jonathan Cohen, Jonathan Bell, Daniel J. Rocke, Russel R. Kahmke, Ahuva Weiss-Meilik, Walter T. Lee, Ricardo Henao, Lawrence Carin, and Shahar Z. Kovalsky. Deep-learning-based screening and ancillary testing for thyroid cytopathology. The Ame...

  3. [11]

    Deep convolutional neural network vgg-16 model for differential diagnosing of papillary thyroid carcinomas in cytological images: a pilot study

    Qing Guan, Yunjun Wang, Bo Ping, Duanshu Li, Jiajun Du, Yu Qin, Hongtao Lu, Xiaochun Wan, and Jun Xiang. Deep convolutional neural network vgg-16 model for differential diagnosing of papillary thyroid carcinomas in cytological images: a pilot study. Journal of Cancer, 10(20):4...

  4. [12]

    An ensemble deep learning for automatic prediction of papillary thyroid carcinoma using fine needle aspiration cytology

    Nguyen Thanh Duc, Yong-Moon Lee, Jae Hyun Park, and Boreom Lee. An ensemble deep learning for automatic prediction of papillary thyroid carcinoma using fine needle aspiration cytology. Expert Systems with Applications, 188:115927, 2022. ISSN 0957-4174

  5. [13]

    Machine-learning-based diagnosis of thyroid fine-needle aspiration biopsy synergistically by papanicolaou staining and refractive index distribution

    Young Ki Lee, Dongmin Ryu, Seungwoo Kim, Juyeon Park, Seog Yun Park, Donghun Ryu, Hayoung Lee, Sungbin Lim, Hyun-Seok Min, YongKeun Park, et al. Machine-learning-based diagnosis of thyroid fine-needle aspiration biopsy synergistically by papanicolaou staining and refractive in...

  6. [14]

    Deep learning-based cell segmentation for rapid optical cytopathology of thyroid cancer

    Peter R Jermain, Martin Oswald, Tenzin Langdun, Santana Wright, Ashraf Khan, Thilo Stadelmann, Ahmed Abdulkadir, and Anna N Yaroslavsky. Deep learning-based cell segmentation for rapid optical cytopathology of thyroid cancer. Scientific Reports, 14(1):16389, 2024

  7. [15]

    Deep learning models for thyroid nodules diagnosis of fine-needle aspiration biopsy: a retrospective, prospective, multicentre study in china

    Jue Wang, Nafen Zheng, Huan Wan, Qinyue Yao, Shijun Jia, Xin Zhang, Sha Fu, Jingliang Ruan, Gui He, Xulin Chen, Suiping Li, Rui Chen, Boan Lai, Jin Wang, Qingping Jiang, Nengtai Ouyang, and Yin Zhang. Deep learning models for thyroid nodules diagnosis of fine-needle aspiration...

  8. [16]

    A study of machine learning models for rapid intraoperative diagnosis of thyroid nodules for clinical practice in china

    Yan Ma, Xiuming Zhang, Zhongliang Yi, Liya Ding, Bojun Cai, Zhinong Jiang, Wangwang Liu, Hong Zou, Xiaomei Wang, and Guoxiang Fu. A study of machine learning models for rapid intraoperative diagnosis of thyroid nodules for clinical practice in china. Cancer Medicine, 13(3):e6854, 2024

  9. [17]

    Artificial intelligence for personalized medicine in thyroid cancer: current status and future perspectives

    Ling-Rui Li, Bo Du, Han-Qing Liu, and Chuang Chen. Artificial intelligence for personalized medicine in thyroid cancer: current status and future perspectives. Frontiers in Oncology, 10:604051, 2021

  10. [18]

    Deep learning based classification of thyroid cancer using different medical imaging modalities: A systematic review

    Maheen Ilyas, Hassaan Malik, Muhammad Adnan, Umair Bashir, Wajahat Anwaar Bukhari, Muhammad Imran Ali Khan, and Adnan Ahmad. Deep learning based classification of thyroid cancer using different medical imaging modalities: A systematic review. VFAST Transactions on Software Eng...

  11. [19]

    Application of artificial intelligence based on deep learning in the diagnosis of thyroid cancer

    Yixing Lu and Bin Zhang. Application of artificial intelligence based on deep learning in the diagnosis of thyroid cancer. Int J Radiat Med Nucl Med, 46(12):760–764, 2022. ISSN 1673-4114

  12. [20]

    The use of artificial intelligence in the diagnosis and classification of thyroid nodules: An update

    Maksymilian Ludwig, Bartlomiej Ludwig, Agnieszka Mikula, Szymon Biernat, Jerzy Rudnicki, and Krzysztof Kaliszewski. The use of artificial intelligence in the diagnosis and classification of thyroid nodules: An update. Cancers, 15(3), 2023. ISSN 2072-6694. 16 A PREPRINT

  13. [21]

    The minefield of indeterminate thyroid nodules: could artificial intelligence be a suitable diagnostic tool? Diagnostic Histopathology, 29(8):396–401, 2023

    Vincenzo Fiorentino, Cristina Pizzimenti, Mariausilia Franchina, Marina Gloria Micali, Fernanda Russotto, Ludovica Pepe, Gaetano Basilio Militi, Pietro Tralongo, Francesco Pierconti, Antonio Ieni, Maurizio Martini, Giovanni Tuccari, Esther Diana Rossi, and Guido Fadda. The min...

  14. [22]

    Ai in thyroid cancer diagnosis: Techniques, trends, and future directions

    Yassine Habchi, Yassine Himeur, Hamza Kheddar, Abdelkrim Boukabou, Shadi Atalla, Ammar Chouchane, Abdelmalik Ouamane, and Wathiq Mansoor. Ai in thyroid cancer diagnosis: Techniques, trends, and future directions. Systems, 11(10), 2023. ISSN 2079-8954

  15. [23]

    Wong, Brie E

    Charles M. Wong, Brie E. Kezlarian, and Oscar Lin. Current status of machine learning in thyroid cytopathology. Journal of Pathology Informatics, 14:100309, 2023. ISSN 2153-3539

  16. [24]

    Applications of machine and deep learning to thyroid cytology and histopathology: a review

    Greg Slabaugh, Luis Beltran, Hasan Rizvi, Panos Deloukas, and Eirini Marouli. Applications of machine and deep learning to thyroid cytology and histopathology: a review. Frontiers in Oncology, 13, 2023. ISSN 2234-943X

  17. [25]

    The application of artificial intelligence to thyroid nodule assessment

    Paola Chiara Rizzo, Stefano Marletta, Nicolò Caldonazzi, Alessia Nottegar, Albino Eccher, Fabio Pagni, Vincenzo L’Imperio, and Liron Pantanowitz. The application of artificial intelligence to thyroid nodule assessment. Diagnostic Histopathology, 30(6):339–343, 2024. ISSN 1756-2317

  18. [26]

    Thyroid nodule characterization: Overview and state of the art of diagnosis with recent developments, from imaging to molecular diagnosis and artificial intelligence

    Emanuele David, Hektor Grazhdani, Giuliana Tattaresu, Alessandra Pittari, Pietro Valerio Foti, Stefano Palmucci, Corrado Spatola, Maria Chiara Lo Greco, Corrado Ini, Francesco Tiralongo, et al. Thyroid nodule characterization: Overview and state of the art of diagnosis with re...

  19. [27]

    Machine learning and vision transformers for thyroid carcinoma diagnosis: A review

    Yassine Habchi, Hamza Kheddar, Yassine Himeur, Abdelkrim Boukabou, Ammar Chouchane, Abdelmalik Ouamane, Shadi Atalla, and Wathiq Mansoor. Machine learning and vision transformers for thyroid carcinoma diagnosis: A review. arXiv preprint arXiv:2403.13843, 2024

  20. [28]

    From data to insights: A comprehensive survey on advanced applications in thyroid cancer research

    Xinyu Zhang, Vincent Lee, and Feng Liu. From data to insights: A comprehensive survey on advanced applications in thyroid cancer research. arXiv preprint arXiv:2401.03722, 2024

  21. [29]

    Use of Machine Learning–Based Software for the Screening of Thyroid Cytopathology Whole Slide Images

    David Dov, Shahar Z Kovalsky, Qizhang Feng, Serge Assaad, Jonathan Cohen, Jonathan Bell, Ricardo Henao, Lawrence Carin, and Danielle Elliott Range. Use of Machine Learning–Based Software for the Screening of Thyroid Cytopathology Whole Slide Images. Archives of Pathology & Lab...

  22. [30]

    Yolov10: Real-time end-to-end object detection

    Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Yolov10: Real-time end-to-end object detection. arXiv preprint arXiv:2405.14458, 2024

  23. [31]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

  24. [32]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

  25. [33]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications

    Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017

  26. [34]

    Densely connected convolutional networks

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017

  27. [35]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105–6114. PMLR, 2019

  28. [36]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

  29. [37]

    Attention is all you need

    A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017

  30. [38]

    Principal component analysis

    Michael Greenacre, Patrick JF Groenen, Trevor Hastie, Alfonso Iodice d’Enza, Angelos Markos, and Elena Tuzhilina. Principal component analysis. Nature Reviews Methods Primers, 2(1):100, 2022

  31. [39]

    Understanding how dimension reduction tools work: an empirical approach to deciphering t-sne, umap, trimap, and pacmap for data visualization

    Yingfan Wang, Haiyang Huang, Cynthia Rudin, and Yaron Shaposhnik. Understanding how dimension reduction tools work: an empirical approach to deciphering t-sne, umap, trimap, and pacmap for data visualization. Journal of Machine Learning Research, 22(201):1–73, 2021

  32. [40]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618–626,...

  33. [41]

    This standardization ensured uniformity in the input size before the data augmentation process

    Initial Resizing: All images were initially resized to a consistent dimension of 1024x768 pixels. This standardization ensured uniformity in the input size before the data augmentation process

  34. [42]

    This second resizing was performed to match the input requirements of the EfficientNetB0 model

    Model Input Resizing: Within the model architecture, the 1024x768 images were further resized to 224x224 pixels. This second resizing was performed to match the input requirements of the EfficientNetB0 model. No other pre-processing steps were performed. The pixel data, after ...

  35. [43]

    Input Decomposition: The original image I is partitioned into a 3 ×4 grid, yielding 13 sub-images I0,I 1,...,I 12 that collectively cover the entire image

  36. [44]

    Token Generation: Each sub-imageIi is processed by a pre-trained classification model to generate a token Ti that encodes local feature information

  37. [45]

    Dimensional Expansion: Each tokenTi is transformed into a higher-dimensional representation via a linear transformation, ensuring uniformity across tokens

  38. [46]

    Transformer Encoder Processing: The sequence of tokens is input inton Transformer encoder blocks (with a multi-head attention mechanism) to refine the feature representations

  39. [47]

    Final Label Prediction: The token corresponding to the primary region (typicallyT0) is passed through a FFNN to produce the final three-dimensional output vector, representing the predicted class labels for the entire image. This grid-based, automated region selection method s...

  40. [48]

    Data Augmentation: Our data augmentation strategy effectively increases the diversity and size of the training set by a factor of 34, mitigating some of the limitations of the original dataset size

  41. [49]

    This architectural choice reduces the risk of overfitting, which is a common concern with smaller datasets

    Model Complexity: We selected a relatively lightweight model (EfficientNetB0) with only 4 million pa- rameters. This architectural choice reduces the risk of overfitting, which is a common concern with smaller datasets

  42. [50]

    External Validation: The strong performance of our model on an independent external validation set (1015 images) provides further evidence of its generalizability and suggests that the model has learned meaningful features from the training data

  43. [51]

    The system has helped me catch a couple of subtle malignant cases that I might have otherwise missed

    Computational Cost: For larger datasets, k-fold cross-validation can become computationally expensive. We considered the trade-off between the benefits of k-fold CV and the associated computational costs, especially in light of the other factors mentioned above (data augmentat...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.