Answers are grounded in this paper and cite their sources.
Thyroid nodule malignancy classification using ultrasound remains challenging owing to the differences in assessment by TI-RADS among different observers as well as the lack of interpretability in deep learning models. This study proposes a framework that is based on hybrid CNN-Transformer architecture with fuzzy feature optimization and visual explanation. This system combines convolution feature extraction, Transformer modeling, fuzzy membership feature selection, Softmax classifier, and Grad-CAM explanation. Model training was done using the TG3K+TN3K dataset, whereas DDTI, AUITD-Algerian, and micro-lesion synthetic dataset were used for testing. The performance evaluation of the model included classification metrics, cross-dataset testing, TI-RADS feature alignment, probability calibration, ablation testing, computation speed, and reader study by three radiologists. The model based on the suggested framework attained 91.8% accuracy, 89.8% sensitivity, 93.0% specificity, and 90.5% F1-score, along with an ROC-AUC of 0.932 and an inference time of 0.54 s per image. The use of Grad-CAM resulted in 82.85% mean feature alignment with TI-RADS region annotations by radiologists. Diagnostic accuracy increased by 9% points with the use of the assistance of AI. The calibration of probability led to a decrease in expected calibration error from 11.6 to 3.8%. However, factors such as disparities between the datasets used, the retrospective manner in which evaluation was done, and the use of artificial perturbations are some of the reasons why there is no immediate application of these findings into practice. This is because further prospective validation from multicenter studies is necessary to confirm clinical applicability of the results.
No takes yet. Share an insight, caveat, or question.
Jingle et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: