Multi-task Learning Framework with Shared Multi-layer Parameters for Segmentation and Classification of Thyroid Cancer Ultrasound Images

With the rising incidence of thyroid cancer, ultrasound imaging has increasingly become a key tool for early diagnosis. Accurate segmentation and classification of thyroid ultrasound images are crucial for enhancing diagnostic accuracy. However, existing studies typically treat segmentation and classification tasks independently, neglecting their complementary nature and limiting diagnostic efficiency and accuracy. In this study, we introduce a multi-task learning architecture that utilizes shared multi-layer parameters, integrating a Hierarchical Shared Network (HSN), a Task Interaction Module (TIM), and a Dynamic Decoding Module (DDM) to jointly optimize both tasks. Thorough experimental assessments using the publicly accessible TN3K and TG3K dataset reveal that our approach attains a classification accuracy of 93.4% and a Dice coefficient of 87.8% for segmentation, significantly outperforming existing methods. Additional ablation experiments confirm the efficacy of the HSN, TIM, and DDM components. The results indicate that the proposed multi-task learning framework effectively enhances the accuracy and efficiency of thyroid cancer diagnosis, providing substantial support for clinical applications.