A DEEP LEARNING FRAMEWORK COMBINING CNN AND GRAPH ATTENTION NETWORKS FOR THYROID ULTRASOUND DIAGNOSIS

Abstract

Thyroid disorders, including hypothyroidism and hyperthyroidism, are widespread endocrine conditions that pose serious health risks worldwide. Early diagnosis is essential to prevent complications, yet conventional diagnostic methods are often constrained by delayed results, dependence on human expertise, and limited accessibility in remote areas. To address these challenges, this study introduces a hybrid deep learning model that integrates Convolutional Neural Networks (CNNs) with Graph Attention Networks (GATs) for automated thyroid disease detection from ultrasound images.

The proposed framework employs EfficientNet-B4 for spatial feature extraction and GAT layers to capture relational dependencies among features, thereby improving classification performance. Using the Algeria Ultrasound Thyroid Dataset (AUTD), the model achieved an accuracy of 92.48%, precision of 93.94%, recall of 92.48%, and an F1-score of 92.87%, significantly outperforming conventional approaches such as Sequential CNN with K-Means clustering (81.5% accuracy). Key contributions include dynamic graph construction for localized feature representation and advanced data augmentation to address class imbalance.

Extensive experiments, confusion matrix evaluations, and multiclass ROC analyses validate the robustness and clinical applicability of the system. This work advances medical AI by offering a scalable, precise, and deployable solution for early thyroid disease detection. Future directions include exploring advanced attention mechanisms and integrating multimodal clinical data to further enhance diagnostic reliability.