Introduction The interpretation of thyroid ultrasonography(USG) is a challenging task for both human and AI. Pre-operative decision making is high-staked and diagnostic AI tools require meaningful information of how a decision was rendered. This study aims to develop and compare performance of multimodal eXplainable AI(XAI) systems for thyroid nodule risk stratification. Methods USG images and clinical data was collected from 227 patients undergoing thyroidectomy. Classification ground truth is exclusively gold-standard surgical histology. The AI architecture was trained to identify the relevant nodule and classify identified nodule or full USG image into benign or malignant. The XAI was subsequently concatenated with clinical data including FNAC results to produce multimodal systems. Gradient-Weighted Class Activation Map(Grad-CAM) is used to provide saliency mapping for visual interpretability of the XAI system’s prediction. Results The XAI systems predicts histology as follows: Model 1 (Segmented nodule + clinical data) Accuracy 77% F-score 70% Sensitivity 61% Specificity 89% AUC 0.87 Model 2 (Full USG image + clinical data) Accuracy 84% F-score 70% Sensitivity 96% Specificity 81% AUC 0.86 Model 3 (Segmented nodule + Full USG image + clinical data) Accuracy 86% F-score 78% Sensitivity 86% Specificity 87% AUC 0.88 Grad-CAM: Maps demonstrate salient areas for a benign nodule diagnosis overlaps spongiform areas and malignant diagnosis salient areas overlap solid components of a partially cystic-solid nodule and microcalcifications within nodules. Conclusion Benchmarking histopathology as ground truth and providing visual interpretability can produce a veritable XAI tool for thyroid nodule diagnostics with risk accountability.