Purpose Accurate preoperative risk stratification of indeterminate thyroid nodules (ITNs) is critical to avoid over or undertreatment. This study developed a Multiclass Classification Model (MCM) using a multimodal artificial intelligence (AI) approach to classify ITNs into papillary thyroid carcinoma (PTC), follicular lesions or benign colloid lesions. Methods The MCM integrated dual-scale ultrasound (USG) imaging inputs - capturing both nodule-specific (localised) and broader USG anatomical (global) features - alongside FNAC results and demographic data, creating a comprehensive diagnostic framework. The MCM was trained on 6000 images, incorporating both institutional and open-source datasets, ensuring heterogeneity in imaging sources. A deep convolutional neural network architecture was used for construction of the MCM. Results The MCM achieved an overall accuracy of 92%, with a sensitivity of 92%, specificity of 96% and an AUC of 0.92. Notably, it distinguished PTC with 95% accuracy and follicular lesions with 93% accuracy. An explainability system was constructed that allowed surgeons to visualise key regions within the USG image influencing the MCM’s diagnosis. Key regions are highlighted in red, with less significant areas shown in cooler tones, resembling the colour gradients used in PET scans. Additionally, confidence intervals are generated by repeated stochastic passes generating multiple predictions for a single USG image which allows for risk assessment. Conclusion This multimodal, dual-scale AI system demonstrates the potential to enhance histological classification of ITNs, guiding personalised therapeutic decisions. Its utility is especially significant in settings where access to genomic testing is limited.