Advanced Multi-Scale Vision Transformer With Neural Architecture Search for Thyroid Cancer Diagnosis Using Ultrasound Images

Accurate and interpretable differentiation of abnormal and suspicious (potentially malignant) thyroid nodules in ultrasound imaging remains a critical yet challenging task, particularly in secondary healthcare settings where rapid triage is essential. Conventional deep learning models, especially convolutional neural networks, often fail to capture long-range contextual dependencies and multi-scale morphological cues (e.g., microcalcifications, irregular margins and heterogeneous echogenicity) that are vital for reliable diagnosis. To address this, we propose a novel Multi-Scale Vision Transformer with Adaptive Feature Fusion (MSViT-AFF), a unified and explainable framework that integrates hierarchical Vision Transformer (ViT) encoders, cross-scale attention, and differentiable Neural Architecture Search (NAS). Our approach automatically learns to weight diagnostically significant scales, emphasizing high-resolution features for microcalcification detection in suspicious nodules, through an end-to-end NAS-optimized fusion mechanism. Evaluated on a rigorously curated and radiologist-validated dataset of 600 thyroid ultrasound images from Sultan Qaboos University Hospital (SQUH), comprising 300 abnormal and 300 suspicious cases, MSViT-AFF achieved 96.2% accuracy, 94.8% sensitivity, 97.1% specificity, and an AUC of 0.983 [95% CI: 0.974, 0.991], exceeding all evaluatedCNNand single-scale transformer baselines with statistically significant margins (allp< 0.01, DeLong’s test). Moreover, a Grad-CAM-based explainability layer provides pixel-level visualizations aligned with established radiological criteria, thereby enhancing clinical trust and transparency. Cross-dataset evaluation with fine-tuning on the public DDTI benchmark yielded an AUC of 0.966, providing encouraging evidence of cross-dataset transferability and generalization potential. With an inference time of only 15 ms per image, the proposed model shows promise for real-time deployment in resource-constrained secondary care units, pending prospective multi-centre validation.