ABSTRACT Thyroid disease diagnosis using medical imaging has greatly benefited from deep learning by enabling automated feature extraction and reducing reliance on expert interpretation. However, challenges such as limited data availability, inadequate model interpretability, and restricted generalization continue to hinder clinical adoption. To address these issues, this research proposes a hybrid framework based on Kolmogorov–Arnold Graph Neural Networks (KAGNN) integrated with the Sobel Sequence‐Based Hiking Optimization Algorithm (SSHOA) for accurate thyroid disease classification and segmentation. KAGNN effectively models spatial dependencies and inter‐feature relationships, while a Dilated Composite Backbone Network (DCBN) extracts multi‐scale contextual features from ultrasound images. SSHOA optimizes KAGNN parameters, ensuring improved convergence and enhanced diagnostic performance. Furthermore, interpretability is strengthened through the integration of a Score‐CAM‐based visualization module, enabling class‐discriminative activation mapping and providing transparent insights into model predictions for clinical validation. Experimental evaluations on benchmark datasets demonstrate superior segmentation with an IoU of 98.12% and a Dice coefficient of 98.97%, and high classification performance with 99.21% accuracy, 98.85% precision, and 99.04% recall, outperforming models such as DW‐Swin, MFMSNet, and GLFNet. Cross‐validation, ablation analysis, and optimization fitness curves further validate the robustness and reliability of the proposed framework. Overall, the combination of KAGNN and SSHOA delivers a powerful, interpretable, and scalable solution for thyroid disease detection, offering substantial improvements in both segmentation and classification for clinical decision support.
Reddy et al. (Fri,) studied this question.