A multi-task deep learning framework for intraoperative diagnosis of thyroid cancer metastasis using whole slide...

Lymph node metastasis (LNM) is a critical prognostic indicator in papillary thyroid carcinoma (PTC), significantly influencing surgical decision-making and the administration of adjuvant therapies. While conventional imaging modalities such as ultrasound, computed tomography, magnetic resonance imaging, and intraoperative frozen sections are widely used, their sensitivity in detecting micrometastases and accurately localising metastatic regions remains limited. Recent advances in deep learning applied to whole slide images (WSIs) offer promise in enhancing diagnostic precision. However, existing approaches are often constrained by suboptimal performance and limited interpretability, hindering clinical integration.

We present a deep learning framework employing Clustering-constrained Attention Multiple Instance Learning (CLAM) to predict three clinically relevant pathological features from intraoperative frozen section WSIs in PTC: (1) presence of lymph node metastasis, (2) T-stage classification, and (3) anatomical localisation of metastatic involvement (central vs. lateral cervical nodes). A dataset comprising 569 patient samples from two independent centres was utilised to assess generalisability. Feature extraction was performed using a set of convolutional neural network architectures, among which ResNet50 demonstrated the highest predictive performance. Model training and evaluation were conducted using a 10-fold Monte Carlo cross-validation strategy. Interpretability was achieved through attention-based Grad-CAM visualisations to localise diagnostically relevant regions within WSIs.

The proposed CLAM-based pipeline achieved robust predictive performance, with area under the receiver operating characteristic curve scores of 0.85 for lymph node metastasis detection, 0.65 for T-stage classification, and 0.71 for anatomical localisation. The model exhibited consistent performance across both institutions. Moreover, incorporating attention-based interpretability mechanisms facilitated meaningful visualisation of relevant tissue regions, supporting potential clinical applicability.

This study introduces a clinically relevant, interpretable, and generalisable deep learning framework for multi-task analysis of intraoperative frozen section WSIs in papillary thyroid carcinoma. By leveraging weakly supervised learning and a diverse multi-centre dataset, the proposed approach demonstrates promise in augmenting intraoperative diagnostics, guiding surgical planning, and advancing precision oncology in thyroid cancer management.