Deep learning-based multimodal radiopathomics for preoperative prediction of lymph node metastasis in papillary thyroid carcinoma

Abstract

Papillary thyroid carcinoma (PTC) exhibits a high incidence and a strong propensity for lymph node metastasis (LNM). Accurate preoperative assessment of LNM is crucial for guiding surgical approaches, but conventional ultrasound exhibits suboptimal sensitivity. In this study, we developed a deep learning-based multimodal model to predict LNM in PTC patients. We retrospectively collected fine needle aspiration (FNA) liquid-based cytology specimens (N = 1095) and corresponding ultrasound images (N = 2190). A ResNet-101 architecture was trained using five-fold cross-validation and validated on external datasets from two independent centers. The multimodal model achieved strong predictive performance on the internal validation set (area under the curve, AUC: 0.891; accuracy: 0.821) and external validation set (AUC: 0.875; accuracy: 0.808). It outperformed models based solely on ultrasound or cytology images. Gradient-weighted class activation mapping revealed that nuclear features in FNA images were the most influential for LNM prediction. Our model achieved promising predictive performance and has the potential to guide clinical decision-making, potentially reducing unnecessary lymph node dissections in PTC patients.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-48693-y.

Introduction

The global incidence of thyroid cancer has risen sharply over the past several decades, making it the seventh most common malignancy worldwide, particularly among women^1,2. Papillary thyroid carcinoma (PTC) accounts for approximately 84% of all thyroid cancers, representing the most prevalent subtype^3,4. Although PTC is generally regarded as an indolent tumor with favorable outcomes, concerns about overdiagnosis and overtreatment have been increasingly recognized^5–7. In particular, lymph node dissection (LND) exposes patients to a higher risk of surgical complications, and the need for accurate preoperative prediction of lymph node metastasis (LNM) is critical^8–11.

Traditional preoperative ultrasound, serving as the primary imaging modality for assessing LNM, offers limited diagnostic value with a sensitivity of merely 20%−30%, thereby guiding surgical planning in only a minority of cases. To address this limitation, numerous studies have explored predictive models for LNM in PTC^12–14. Early approaches primarily relied on statistical models constructed based on clinical risk factors^15,16. Given that nodule size is an independent risk factor for LNM—particularly for nodules larger than 1 cm—we restricted our study to nodules ≤ 2 cm to mitigate potential bias from significant variations in nodule size and to better elucidate factors beyond size in predicting LNM. Recent studies have extracted high-throughput radiomics features from ultrasound images, confirming their correlation with LNM status to construct robust predictive models^17,18.

Despite these advances, the predictive performance of these models remains suboptimal, suggesting that applying deep learning may offer a potential pathway to address this shortcoming. Deep learning algorithms have demonstrated remarkable efficacy in medical image analysis, spanning fields such as histopathology and cytology^19,20. In deep neural network (DNN) architectures, convolutional neural networks (CNNs) have gained widespread application in medical imaging in recent years due to their particular suitability for image classification tasks. With significant advancements in artificial intelligence (AI), CNNs have further spurred deep learning applications at the multimodal data level^21,22, aiming to enhance predictive accuracy by integrating complementary data from multiple sources^23,24. For example, in thyroid cancer, cytological features from fine-needle aspiration (FNA), such as nuclear enlargement, crowding, overlapping, longitudinal grooves, nuclear pseudoinclusions, and ground-glass chromatin, provide valuable microscopic information that complements ultrasound-derived macroscopic features, including tumor contour, margin characteristics, and echogenicity. This multimodal strategy offers a more comprehensive framework for predicting tumor aggressiveness and metastatic potential.

In this study, we developed an integrated radiopathomics model that combines ultrasound and FNA images through DNN and a novel image patch processing strategy. By embedding and jointly analyzing preoperative multimodal imaging data, our model enables robust prediction of LNM in PTC, with the potential to guide surgical decision-making and reduce unnecessary LND procedures.

Results

Baseline characteristics

From the initial cohort of 1,295 patients, the internal validation set comprised 1,095 patients (median age: 42 years; range: 19–83 years), yielding 1,095 whole slide images (WSIs) and 2,190 paired ultrasound images sourced from The First Affiliated Hospital of the Medical School of Zhejiang University for model development. Among these patients, 75% were female. In the non-metastatic group, the proportion of female patients increased to 82%, consistent with prior reports indicating a higher prevalence of PTC in women^1,2. According to the ultrasound thyroid imaging reporting and data system (TI-RADS) classification, most patients were categorized as 4a (59%) and 4b (31%). The distribution across the two groups (N0, N1) showed no significant differences, suggesting that TI-RADS classification is not strongly associated with LNM. Baseline characteristics are summarized in Fig. 1; Table 1, and there’s no statistically significant differences in baseline characteristics between the internal cohort and external cohort (Table 1 and Supplementary Table S1).

Fig. 1.

Baseline characteristics of the study cohort.

Table 1.

Baseline characteristics of patients.

Characteristics	N0	N1	Internal overall	External overall	P value (Internal v external)
No. patients	558	537	1095	200	/
Age (y)	44 (24–73)	39 (19–83)	42 (19–83)	41 (21–76)	0.128
Female	455 (82%)	368 (70%)	821 (75%)	156 (78%)	0.356
TI-RADS
4a	357 (64%)	287 (56%)	644 (59%)	115 (58%)	0.292
4b	147 (26%)	194 (35%)	341 (31%)	65 (32%)	0.087
4c	18 (4%)	26 (4%)	44 (4%)	10 (5%)	0.241
5	36 (6%)	30 (5%)	66 (6%)	10 (5%)	0.095

Note: Age were presented as median (range). Other data were presented as number and percentages. TI-RADS, thyroid imaging reporting and data system. P-values were obtained using the independent samples t-test.

Prediction performance of deep learning models

The single-modality model based solely on ultrasound images demonstrated suboptimal predictive performance. In the internal validation cohort (Fig. 2 (a), (b), blue curves), the area under the curve (AUC) was 0.759, and the area under the precision-recall curve (AUCPR) was 0.752. In the external validation cohort, the AUC and AUCPR were 0.744 and 0.742, respectively, all below the threshold of 0.80 (Fig. 2, (d), (e), blue curves). In the single-modality deep learning model trained exclusively on FNA images, predictive performance improved compared with the ultrasound-only model. In the internal validation cohort, the model achieved an AUC of 0.851 and an AUCPR of 0.820 (Fig. 2 (a), (b), green curves). In the external validation cohort, the AUC and AUCPR were 0.820 and 0.810, respectively, indicating moderate generalizability but leaving room for further optimization (Fig. 2 (d), (e), green curves).

Fig. 2.

Performance of LNM prediction in fine-needle aspiration alone (FNA) model, ultrasonography alone (US) model and multimodal radiopathomics model (FNA-US). (a), (d) ROC curves; (b), (e) Precision-Recall curves; (c), (f) Confusion matrices. Within the confusion matrices, the top-left quadrant denotes true negatives (correctly predicting the absence of metastasis), whereas the top-right quadrant indicates false negatives (predicting no metastasis despite actual presence).,the bottom-left grid represents cases predicted as metastasis with actual no metastasis (false positive rate), and the bottom-right grid represents cases predicted as metastasis with actual metastasis (true positive rate).Panels (a)-(c): internal validation set; panels (d)-(f): external validation set.

The multimodal radiopathomics model, which integrates ultrasound and FNA images through feature-level fusion, utilized complementary information from both modalities and markedly enhanced diagnostic performance. In comparison to single-modality approaches, the multimodal framework consistently demonstrated superior predictive accuracy in both internal and external validation cohorts. In the internal validation set (Fig. 2, (a), (b), purple curves), the multimodal model achieved an AUC of 0.891, representing relative improvements of 5.1% and 17.4% over the FNA-only and ultrasound-only models, respectively. The AUCPR reached 0.852, corresponding to increases of 4.1% and 13.3% compared with the FNA and ultrasound single-modality models, respectively. These findings suggest that combining structural features from US images (e.g., nodule margins and blood flow distribution) with the cellular morphological features from FNA images (e.g., nucleocytoplasmic ratio and cellular atypia) enables the model to capture richer cross-modal feature associations, thereby mitigating diagnostic bias inherent to single-modality approaches. In the external multi-center validation set (Fig. 2, (d), (e), purple curves), the multimodal model achieved an AUC of 0.875 and an AUCPR of 0.840. These values were substantially higher than those of single-modality models (AUC increased by 6.7% and AUCPR by 3.8% compared with the FNA model; AUC increased by 17.6% and AUCPR by 13.2% compared with the ultrasound model). Confusion matrices and balanced accuracy values are shown in Fig. 2 (c), (f), while additional multimodal model characteristics are summarized in Fig. 3. and Table S2. Overall, the multimodal radiopathomics model demonstrated robust and clinically promising performance.

Fig. 3.

Model performance in the internal and external validation cohorts. Accuracy (Cyan) describes the proportion of correctly classified cases among the total number of cases assessed. Precision (Orange) describes the proportion of true positive predictions among all cases predicted as positive. Recall (Blue) describes the proportion of actual positives correctly identified. F1-Score (Gray) describes the harmonic mean of Precision and Recall, providing a single metric that balances both concerns.

Model explainability

Representative cases are shown in Fig. 4, illustrating regions of interest highlighted in FNA images and ultrasound images. Visualization results indicated that the model primarily focused on clustered nuclear regions in FNA images (Fig. 4, (a), colored in red), while stromal, vascular, and blank areas were assigned low weights (Fig. 4, (a), colored in blue or green), reflecting minimal importance. Regarding the ultrasound modality, gradient-weighted class activation mapping (Grad-CAM) heatmaps revealed a primary model focus on calcification regions (Fig. 4, (b), colored in red). Furthermore, Spearman correlation analysis revealed moderate complementarity between the feature spaces of the two modalities (ρ = 0.32, P < 0.01), supporting the biological plausibility and scientific validity of the radiopathomics fusion strategy.

Fig. 4.

Visualization of representative LNM-positive cases in the validation set, with Grad-CAM heatmaps highlighting modality-specific regions of interest. (a) displays the original fine-needle aspiration cytology images alongside their corresponding Grad-CAM heatmaps; (b) presents the original ultrasound images and their associated Grad-CAM heatmaps. Grad-CAM, gradient-weighted class activation mapping.

Discussion

Our study developed a deep learning-based multimodal framework that integrates preoperative ultrasound and FNA cytology images to predict LNM in PTC. The model demonstrated robust performance, achieving an AUC of 0.891 in the internal validation set and 0.875 in the external multicentre validation set, outperforming models utilising either ultrasound or FNA images alone. This underscores the synergistic effect of combining macroscopic sonographic features with microscopic cytological characteristics, providing a more comprehensive assessment of tumour aggressiveness. The superior generalisability of our multimodal approach is evident from the relatively minor performance degradation observed in external validation compared to single-modality models, highlighting its potential for broader clinical application.

The findings of this study align with the growing body of evidence supporting the value of multimodal AI in medical image analysis. Recent research has similarly demonstrated that integrating complementary data sources enhances predictive accuracy for LNM in thyroid cancer. For instance, studies have reported successful fusion of ultrasound radiomics with deep learning features^25–32, computed tomography (CT) images with iodine maps, and magnetic resonance imaging (MRI) data with clinicopathological information^29,30. Our model contributes to this emerging paradigm by specifically focusing on the fusion of routinely available preoperative imaging modalities—ultrasound and FNA cytology—which are particularly relevant for clinical decision-making in PTC.

When contextualised within the current literature, our findings both confirm and extend previous work. Previous studies have primarily focused on identifying image patches containing follicular cell clusters with discriminative features for PTC, followed by classification^33,34. Although effective in some cases, such methods have inherent limitations. First, reliance on follicular cell clusters reduces applicability, as some WSIs contain a few clusters or none at all. Our method addresses this by incorporating both isolated nuclei and cell clusters, thereby extending applicability across diverse image scenarios. Second, conventional approaches typically require labor-intensive manual annotation of regions of interest by pathologists. In contrast, our framework requires only slide-level labels for training, substantially reducing annotation burden and improving scalability. Third, patch-based WSIs segmentation generates large numbers of inputs for neural networks, resulting in high computational costs. Our approach alleviates this burden, achieving superior efficiency in both computational resource utilization and model performance. The superior performance of our multimodal model is consistent with studies reporting that AI-based fusion models outperform single-modality approaches for predicting LNM in PTC. However, our model achieves this with a focus on the most commonly available preoperative data—ultrasound and FNA images—enhancing its potential for practical implementation. The external validation across multiple centres strengthens the evidence for its generalisability, a critical consideration for real-world deployment.

Additionally, to mitigate the “black‑box” limitation of deep learning-based models, Grad-CAM was incorporated to generate attention heatmaps highlighting macro- and micro-level features most influential in the prediction. These visualizations identified key ultrasound regions and cytological patterns, enhancing transparency and supporting clinical applicability. The heatmaps indicated that the model primarily attended to calcification areas in ultrasound images and clustered nuclear regions in FNA images, which is biologically plausible given that these features are clinically associated with tumour aggressiveness and metastatic potential. This focus on interpretability addresses a significant barrier to clinical adoption of AI systems, as understanding the rationale behind predictions is crucial for building trust among clinicians. The moderate complementarity observed between the feature spaces of the two modalities (Spearman correlation ρ = 0.32) further supports the validity of our fusion strategy.

While existing prediction models for LNM in PTC utilize ultrasound, CT, or MRI^35–37, our framework offers distinct advantages in modality selection, clinical applicability, and health economics. Our methodology achieves unprecedented deep learning fusion between macroscopic ultrasound features and microscopic whole-slide cytopathology images, effectively bridging the inherent pixel gap. Whereas the existing multimodal model of ultrasound and CT/MRI images falls within the integration of macroscopic anatomical imaging, our model enables the fusion of macroscopic ultrasound features (e.g., tumor margins, echogenicity, and spatial relationships with adjacent tissues) with cytomorphologic features (e.g., nuclear crowding, grooves, and intranuclear pseudoinclusions) within a unified deep learning framework. By combining macro-structural and micro-morphological data, the model leverages complementary information across modalities. In the validation cohort, it achieved an AUC of 0.891, an AUCPR of 0.852, and an accuracy of 0.821, representing relative improvements of 5.1%, 17.4%, and 4.1%, respectively, over single-modality baselines. Secondly, with respect to direct feasibility for clinical application, our model holds a distinct advantage. Although MRI-based models may provide rich diagnostic information, it is not routinely employed as a preoperative examination for thyroid cancer and would significantly increase costs and healthcare burden. In contrast, the ultrasound and FNA cytology procedures underpinning our model are “gold standard” and preoperative tests must be performed. Consequently, our model can be directly integrated into existing clinical pathways without requiring any alterations, demonstrating superior “ready-to-use” practicality compared to models dependent on CT or MRI. Economically, compared to resource-intensive CT or MRI approaches, our model relies exclusively on routine ultrasound and FNA cytology. This strategy minimizes patient financial burden and contrast-associated risks, maximizing clinical accessibility.

Despite these promising results, several limitations warrant consideration. The retrospective study design and the inclusion criteria requiring adequate cell quantity (excluding samples with low cell count) may introduce selection bias and limit the generalizability of our findings. Although external validation was performed, the high standardization of protocols and equipment across centers may overestimate the model’s robustness to real-world technical variations; prospective studies in more diverse, multi-institutional settings are needed. Furthermore, the manual annotation of ultrasound ROIs, though standardized, introduces a source of inter-observer variability. Finally, our multimodal fusion aligns data at the patient level, which is clinically pragmatic but does not enforce precise spatial registration between ultrasound images and cytology patches, potentially not capturing fine-grained spatial correspondences.

Thus, future research should focus on integrating additional data modalities, such as clinical biomarkers (e.g., biomarkers mutation status) or other imaging features, to further enhance predictive accuracy. Efforts to develop more automated feature extraction pipelines and more sophisticated fusion algorithms could also improve efficiency and performance. Most importantly, prospective clinical validation is essential to demonstrate the impact of this tool on patient outcomes and surgical decision-making.

In summary, our deep learning-based multimodal model provides a reference framework in the preoperative prediction of LNM in PTC, outperforming single-modality approaches and demonstrating good generalizability. By leveraging complementary information from ultrasound and FNA cytology, the model provides a more comprehensive assessment of metastatic risk, potentially guiding more individualized treatment strategies. The incorporation of explainable AI techniques enhances its clinical translatability. Future work should focus on prospective validation and integration of additional data sources to further refine its predictive capability.

Methods

Ethics statement

All methods were performed in accordance with the relevant guidelines and regulations. This retrospective study was approved by the Ethics Review Committee of the First Affiliated Hospital, Zhejiang University School of Medicine (Approval No. ⅡT20240373A). This committee specifically waived the requirement for informed consent due to the retrospective nature of the study.

Data collection

Patients with pathologically confirmed PTC who underwent thyroidectomy and central neck lymph node dissection at the First Affiliated Hospital, Zhejiang University between 2021 and 2023 were included. Inclusion criteria: (1) Availability of preoperative ultrasound and FNA; (2) Complete clinical data; (3) Age ≥ 18 years. (4) The greatest dimension ≤ 2 cm. Exclusion criteria: (1) Prior preoperative ablation, radiotherapy, or chemotherapy; (2) Multiple thyroid nodules with concurrent PTC; (3) Incomplete imaging data; (4) Age < 18 years. For external validation, additional data (200 samples, between 2021 and 2023) were collected from Zhejiang Provincial People’s Hospital and Ningbo Pathology Diagnosis Center. The two additional centers utilized identical protocols to the primary center for the cytological preparation of FNA samples, image scanning hardware, and image acquisition parameters. Similarly, the ultrasound image acquisition equipment and patch acquisition methodology were consistent with the primary center and the same three ultrasound doctors interpreted the images using the same methodology (Fig. S1).

Liquid-based cytology (LBC) WSIs

A total of 1,295 FNA LBC WSIs were retrieved from the pathological sample archive (1095 WSIs for internal training and validation, 200 WSIs for external validation). Following procurement, all FNA liquid-based cytology specimens underwent prompt processing, standard Papanicolaou staining, and neutral synthetic resin mounting. Two independent pathologists possessing over five years of diagnostic experience (Attending Physicians A and B) initially evaluated the slides, followed by a consensus review by a senior pathologist with over a decade of expertise (Associate Chief Physician C). All three physicians were blinded to the patients’ clinical data. Slides with sufficient cellularity (containing at least six clusters of follicular epithelial cells, with each cluster comprising 10–20 cells³⁸) were included, whereas poor-quality slides (scant cellularity, faded staining, blood contamination, staining artifacts) were excluded. For all samples, postoperative histopathology confirmed the lymph node status (N0 or N1). All qualified slides were scanned at 20× magnification (0.5 μm/pixel) using two high-performance slide scanners (KF-PRO-400, Ningbo Konfoong Biotech International Co., Ltd.) and saved in KFB file format.

Ultrasound images

Three experienced ultrasound physicians conducted all sonographic examinations utilizing high-resolution systems using high-resolution systems (Philips iU22 or GE Logiq E9, 7–12 MHz linear probe). Ultrasound images of 1295 patients were collected (1095 patients for internal training and validation, 200 patients for external validation), nodule morphology, margin definition, and internal echogenic features were documented for model development

Artificial intelligence model

An overview of the proposed algorithm is shown in Fig. 5. Pathological WSIs were segmented into 64 × 64-pixel patches, with regions containing isolated nuclei (300–2000 pixels, approximately 75–500 µm² in area) and clustered cellular structures such as papillary formations and psammoma bodies retained for analysis. A ResNet-101 CNNs was employed to classify valid nuclei patches while filtering out artifacts, impurities, and blurred regions. For ultrasound images, target nodule regions were manually cropped by physicians, rescaled to a fixed resolution, and intensity-standardized to ensure the preservation of key morphological and echogenic features, including nodule size, margin clarity, and blood flow characteristics. Importantly, cross-validation was conducted with patient-level separation. All patches and images from the same patient were assigned to a single fold, and model training followed this split to prevent data leakage.

Fig. 5.

Overview of the deep learning model pipeline. (A) 5-fold cross-validation split of pathology WSIs for N0 and N1ab. (B) Pathology and ultrasound feature extraction and fusion into a combined sequence. (C) Attention-based fusion to produce a fused global token for prediction.

Deep learning feature selection

Nuclei in malignant PTC cytology slides often display distinctive morphological features, such as nuclear grooves and nuclear inclusions, which may serve as indicators of LNM risk. To isolate individual nuclei, only masked regions with areas between 300 and 2000 pixels were retained. From these regions, 64 × 64-pixel patches were extracted from WSIs, with the reserved nuclei regions placed at the centers (Fig. 6). Specifically, we scanned each WSI at level 0 using non-overlapping 4096 × 4096 tiles within a central circular region, detected nuclei in each tile, and generated one patch per qualified nucleus. As a result, the number of patches per WSI varied with nuclear yield and ranged from 47 to 2867, with a mean of 1020. Although most patches contained intact, well-focused nuclei, some included artifacts, impurities, or blurred nuclei. To address this, CNNs was employed to classify patches into valid nuclear images versus non-informative (impurity) patches, and only valid patches after CNN-based quality filtering were retained for subsequent modeling. In addition to isolated nuclei, clustered cell groups also exhibit diagnostically relevant features, such as papillary structures, psammoma bodies, and disturbed nuclear polarity. Therefore, all identifiable cell clusters in WSIs were collected to enable feature extraction for model input. For multimodal prediction, pathology patch features from each WSI were treated as a token sequence and aggregated by multi-head attention blocks to obtain a slide-level pathology representation, ultrasound images were encoded into an ultrasound feature sequence, and the two modalities were fused by a two-way attention-based encoder into a joint representation for final classification. The two modality-specific sequences were concatenated and iteratively updated through token self-attention and two-way cross-attention to capture bidirectional interactions. The final prediction was produced from the fused global token embedding.

Fig. 6.

Procedure for removing blank regions.

During training, neural networks typically prioritize “hard” samples with higher loss values; however, these are often noisy or mislabeled samples. To mitigate this issue, we introduced an adaptive label-switching strategy controlled by a probability parameter , which increases with training epochs. This adjustment facilitates correction of mislabeled or noisy samples, allowing the model to learn more reliable representations. The probability is defined as follows:

where E denotes the total number of training epochs, e denotes the current epoch, and q is a hyperparameter controlling the rate of increase.

As for Ultrasound images, the eligible ultrasound images from PTC patients were exported in DICOM format and uploaded to the Radcloud research platform. A multi-step annotation workflow was then performed: two ultrasound doctors with over five years of thyroid diagnosis experience (Attending Physicians A and B) independently delineated regions of interest (ROIs) along the margins of the thyroid cancer foci on both longitudinal and transverse planes. Subsequently, all contours were reviewed by an ultrasound doctor with over ten years of thyroid diagnosis experience (Associate Chief Physician C). All three physicians were blinded to the patients’ clinical data and pathological results of the target thyroid nodules. The eligible ultrasound image patches were included for model development.

Prediction model based on extracted features

A deep learning framework was employed to capture cytological heterogeneity and encode this information into high-dimensional feature vectors. For each patient, the cytology WSI was paired with ultrasound images of the same target thyroid nodule acquired in the transverse and longitudinal planes. All ultrasound images for that patient were grouped at the patient level and encoded into an ultrasound feature sequence, while the cytology WSI was encoded into a pathology feature sequence from the extracted patches. This design aligns the two modalities at the patient and nodule level rather than enforcing one-to-one matching between individual ultrasound images and pathology patches. To address the imbalance in image counts across modalities, fusion was performed on variable-length sequences through attention, and the ultrasound sequence length was capped by uniform sampling during training when needed. The extracted features were concatenated and input into a bidirectional attention network to model cross-modal interactions between ultrasound and pathological data. Subsequently, a final attention module refined the fused representations, and a multi-layer perceptron was used to predict N0/N1ab metastasis.

Model reliability and explainability

To improve clinical interpretability, model explainability was assessed using Grad-CAM. This technique highlights image regions that contribute most to the prediction by computing gradients of the class score with respect to convolutional feature maps. Formally, consider a 2D image classification network producing a score for class . For a convolutional layer with k feature maps (channels), , where and denote spatial indices, the neuron importance weight is defined as:

where n is the total number of pixels in the feature map. The Grad-CAM heatmap is obtained as a weighted combination of these feature maps and, when superimposed on the original image, provides a visual explanation of the model’s decision process.

Computation and statistical analysis

Deep learning model training and feature extraction were executed utilizing Python 3.10.14 and PyTorch 2.7.0 with CUDA 12.4 on an NVIDIA RTX A6000 GPU. Subsequent statistical evaluations were conducted in MATLAB 2023a on an Intel Core i7 processor. Cross-validation runtimes were linear discriminant analysis (1.74–5.54 s/run) vs. support vector machine (22.15–193.24 s/run). Model performance was quantified using balanced accuracy, area under the receiver operating characteristic curve, and AUCPR, ensuring reliable evaluation across datasets.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

LNM

Lymph node metastasis

PTC

Papillary thyroid carcinoma

FNA

Fine needle aspiration

Grad-CAM

Gradient-weighted class activation mapping

LND

Lymph node dissection

DNN

Deep neural network

CNNs

Convolutional neural networks

AI

Artificial intelligence

WSIs

Whole slide images

AUC

Area under the curve

AUCPR

Area under the precision-recall curve

LBC

Liquid-based cytology

CT

Computed tomography

MRI

Magnetic resonance imaging

Author contributions

Conceptualization: Qihan You, Qiuyu Cai, Jing Zhang. Methodology: Tianhong Gao, Zunlei Feng. Software: Tianhong Gao, Zunlei Feng. Validation: Qihan You, Qiuyu Cai, Tianhong Gao, Jing Zhang. Formal analysis: Tianhong Gao, Zunlei Feng. Investigation: Xianfa Xu, Yi Chen, Tian-An Jiang, Qiuyu Cai. Resources: Xianfa Xu, Yi Chen, Tian-An Jiang, Qiuyu Cai. Data Curation: Tianhong Gao, Linyun Zhou, Shengxuming Zhang. Writing - Original Draft: Qiuyu Cai, Tianhong Gao. Writing - Review & Editing: Qihan You, Xiuming Zhang, Jing Zhang. Visualization: Tianhong Gao, Zunlei Feng. Supervision: Qihan You, Xiuming Zhang, Jing Zhang. Project administration: Qihan You, Jing Zhang, Zunlei Feng. Funding acquisition: Xiuming Zhang.

Funding

This work was supported by the National Natural Science Foundation of China (Grant/Award Number: 62376248, Zunlei Feng) and Natural Science Foundation of Zhejiang Province (Grant/Award Number: LQ20H160048 and LY21H160035, Xiuming Zhang).

Data availability

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.