CLNAD-Net: A Multi-Task Learning Framework for Cervical Lymph Node Computer-Aided Diagnosis Network.

Authors

Abstract

The accurate diagnosis of cervical lymph node metastasis (CLNM) is critical for selecting appropriate treatment plans for patients with papillary thyroid cancer (PTC). Currently, diagnostic processes primarily rely on the expertise of ultrasound physicians. However, challenges such as low contrast and high noise in ultrasound images often lead to variability in diagnostic outcomes, particularly when performed by less experienced physicians or under conditions of fatigue. To address these limitations, this study proposes CLNAD-Net, a multi-task learning network designed to simultaneously perform automatic segmentation and metastatic classification of cervical lymph node ultrasound images. CLNAD-Net employs a dual-branch encoder architecture that integrates the strengths of CNNs and Transformer models. Through its feature alignment module (FAM), the network effectively fuses local and global features, while the multi-modal auxiliary classification module (MACM) enhances classification tasks by incorporating both ultrasound images and patient clinical data to improve diagnostic accuracy. The study utilized a dataset comprising 541 cervical lymph node ultrasound images from 460 patients, collected at the Chenzhou No.1 People's Hospital. Experimental results demonstrate that CLNAD-Net outperformed existing methods and physicians of varying experience levels in both segmentation and classification tasks, achieving an accuracy of 84.49% in classification. Moreover, the inclusion of clinical data improved classification accuracy by 1.78%. These findings highlight the significant potential of CLNAD-Net in CLNM diagnosis, offering improved efficiency and accuracy for clinical applications. The proposed method provides a promising framework for the development of automated auxiliary diagnostic systems in medical imaging.

Tags

Topics