Evaluating the Diagnostic and Explainability Impact of Adding TIRADS Information to Deep Learning for Thyroid Nodule Malignancy Classification: A Comparative Study Across Two Ultrasound Datasets

Background: Hybrid ultrasound AI models can combine image features with structured reporting information, but it is important to investigate whether the model attends to clinically relevant image regions. This study evaluated whether adding TI-RADS-related information to image-based deep learning changes both thyroid nodule malignancy classification and Grad-CAM-based attention to segmented nodule regions. Methods: Image-only and hybrid EfficientNet-B0 models were compared across two public thyroid ultrasound datasets: the Stanford AIMI Thyroid Ultrasound Cine-clip dataset and ThyroidXL. The Stanford hybrid model fused ultrasound frames with ACR TI-RADS descriptor fields, while ThyroidXL hybrid models fused images with either nodule position/size metadata or final TI-RADS category. Diagnostic performance was evaluated at the patient level using AUC, accuracy, F1 score, sensitivity, and specificity. Explainability was evaluated by comparing Grad-CAM heatmaps with segmentation masks using overlap-based metrics including IoU, Dice coefficient, and inside-mask activation. Results: Adding TI-RADS-related information improved diagnostic performance in both datasets. In Stanford, AUC increased from 0.704 for the image-only model to 0.944 for the hybrid model, although this result was limited by a small test set with only two malignant patients. In ThyroidXL, AUC increased from 0.921 for the image-only model to 0.929 with nodule metadata and 0.965 with final TI-RADS category fusion. Explainability results were not consistent across datasets. In a selected Stanford subset, the hybrid model showed lower Grad-CAM/mask overlap than the image-only model. In the full ThyroidXL test set, the image-plus-final-TI-RADS model showed higher mean IoU, Dice coefficient, and inside-mask activation, with the improvement driven mainly by malignant nodules. Conclusion: TI-RADS-related structured information improved thyroid nodule malignancy classification, but its effect on model attention was dataset and model-dependent. These findings suggest that hybrid ultrasound AI models should be evaluated by diagnostic accuracy, along with their visual attention to nodule regions.