UWMT-net: An uncertainty-weighted multi-task network for ultrasound-based three-class classification of thyroid nodules.
Authors
Affiliations (5)
Affiliations (5)
- Department of Medical Ultrasound, The First Affiliated Hospital of Shandong First Medical University & Shandong Provincial Qianfoshan Hospital, Jinan, Shandong, 250014, China.
- School of Control Science and Engineering, Shandong University, Jinan, Shandong, 250061, China.
- Department of General Surgery, The First Affiliated Hospital of Shandong First Medical University & Shandong Provincial Qianfoshan Hospital, Jinan, Shandong, 250014, China.
- Department of Urology, The First Affiliated Hospital of Shandong First Medical University & Shandong Provincial Qianfoshan Hospital, Jinan, 250014, China. Electronic address: [email protected].
- Department of Urology, The First Affiliated Hospital of Shandong First Medical University & Shandong Provincial Qianfoshan Hospital, Jinan, 250014, China. Electronic address: [email protected].
Abstract
Risk stratification of thyroid nodules on ultrasound remains inconsistent in routine practice, particularly for borderline presentations. Many deep-learning studies report discrimination alone, while probability reliability and threshold-based clinical utility receive less attention. We retrospectively included 364 development patients, each contributing one thyroid nodule labeled using surgical histopathology or fine-needle aspiration cytology. The primary analysis used a patient-level fixed training/validation/test framework, and robustness was additionally assessed with stratified patient-level five-fold cross-validation. A locked model was independently evaluated in 121 patients from the Affiliated Hospital of Shandong University of Traditional Chinese Medicine (Shandong Provincial Hospital of Traditional Chinese Medicine), Jinan, China. The task was formulated as three-class classification: benign, indeterminate or suspicious, and malignant. The multimodal model outperformed clinical-only and image-only baselines, with the largest gains observed on class-balanced metrics. On the held-out test set, UWMT-Net achieved an accuracy of 0.838 and a macro-AUC of 0.943. Post-hoc temperature scaling reduced ECE from 0.040 to 0.015 without changing hard-label metrics. Five-fold cross-validation yielded a macro-AUC of 0.936 ± 0.011. In the independent external cohort, the locked model achieved an accuracy of 0.810, macro-F1 of 0.804, macro-AUC of 0.921, and ECE of 0.031. The added fusion baselines and paired bootstrap analysis supported a modest, but not statistically significant, numerical advantage of uncertainty-weighted multi-task learning. Integrating ultrasound with structured clinical information improved three-class classification and probability reliability in the development cohort, with retained performance in the independent external cohort. Larger prospective multicenter validation remains required before clinical use.