IMPROVING THYROID ULTRASOUND DIAGNOSIS USING A HYBRID CNN-GAT MODEL

Abstract

Thyroid disorders like hypothyroidism and hyperthyroidism are widespread and require early diagnosis to avoid serious complications. Traditional diagnostic methods often suffer from delays, expert dependence, and limited access in remote areas. This study proposes a hybrid deep learning model combining EfficientNet-B4 for spatial feature extraction and Graph Attention Networks (GATs) for capturing feature relationships in thyroid ultrasound images. Trained on the Algeria Ultrasound Thyroid Dataset (AUTD), the model achieved 92.48% accuracy, 93.94% precision, 92.48% recall, and a 92.87% F1-score, outperforming traditional methods such as CNN with K-Means clustering. Key innovations include dynamic graph construction and advanced data augmentation to address class imbalance. Comprehensive evaluation through confusion matrices and ROC curves supports its clinical reliability. The approach offers a scalable and accurate solution for early thyroid disease detection, with future work focusing on improved attention mechanisms and integration with broader clinical data.