Hybrid-supervised multi-branch neural network with semantic perception for thyroid nodule diagnosis using ultrasound images

Thyroid cancer incidence has increased steadily worldwide, making early and reliable diagnosis of thyroid nodules an important clinical task. Ultrasound is the primary imaging modality for thyroid nodule assessment, but manual interpretation remains operator-dependent and is affected by inter-observer variability. In recent years, artificial intelligence (AI)-assisted diagnosis has shown promising progress in thyroid ultrasound image analysis. However, many existing methods still rely mainly on image-level benign-malignant classification and lack clinically structured intermediate representations, which limits model transparency and robustness. In this study, we propose a hybrid-supervised multi-branch neural network for ultrasound-based thyroid nodule diagnosis. The proposed model introduces Thyroid Imaging Reporting and Data System (TI-RADS)-guided semantic perception as an intermediate layer between ultrasound image representation and malignancy prediction, and jointly uses attribute-level semantic supervision, response-distribution-based structural regularization, and final benign-malignant classification supervision. The multi-branch semantic perception module learns clinically meaningful attributes, including composition, echogenicity, shape, margin, and calcification, while a global diagnostic branch captures complementary image-level diagnostic cues beyond predefined attributes. A prototype-guided semantic alignment and fusion module further integrates structured semantic representations and global diagnostic representations. Experimental results show that the proposed model achieves competitive overall diagnostic performance compared with single-task classification models, generic multi-task models, and thyroid-specific diagnostic frameworks. This study provides a clinically structured AI approach for ultrasound-based thyroid nodule diagnosis.