Thyroid pathology image classification via multi-scale feature fusion and multi-instance learning.

BACKGROUND

The global incidence of thyroid cancer has significantly increased, while traditional pathological diagnosis remains time-consuming and expert-dependent. This study develops an auxiliary diagnostic tool designed to reduce the workload of pathologists and improve diagnostic accuracy.

METHODS

Our study utilized 543 WSIs from Liuzhou Cancer Hospital for model development, employing a novel multi-feature fusion architecture that combines RetCCL, iBOT, and DINO embeddings. We systematically evaluated stain normalization and multi-scale analysis across four multiple-instance learning (MIL) frameworks: CLAM-SB (single-branch), CLAM-MB (multi-branch), DTFD (double-tier), and LA-MIL (location-aware). The method was rigorously validated on an independent set of 128 WSIs from Taizhou Cancer Hospital.

RESULTS

The results show that stain normalization, multi-scale fusion, and multi-feature fusion significantly improve classification performance. In 10-fold cross-validation on the internal dataset, the system demonstrated significant improvements over the baseline RetCCL model: AUC (0.9900 vs. 0.9629), accuracy (0.9594 vs. 0.8951), with relative improvements of 2.8% in AUC and 7.2% in accuracy. Precision increased by 11.5% (0.9434 vs. 0.8461) and F1-score by 9.8% (0.9511 vs. 0.8665). On the external validation dataset, the model maintained robust performance with an AUC of 0.9584, accuracy of 0.9070, precision of 0.9247, and F1-score of 0.9348, confirming its reliability and applicability.

CONCLUSIONS

We propose a weakly supervised MIL framework integrating multi-scale analysis and cross-model feature fusion for thyroid cancer diagnosis. Our method showed promising and consistent results across internal and external datasets. While further clinical validation and workflow integration are needed, the results suggest the potential of this approach to assist pathologists in diagnostic workflows, particularly in resource-constrained settings.