EyePterNet-X: A Cross-Attention Hybrid CNN-Transformer Framework for Automated Pterygium Analysis Using Anterior Segment Eye Images

Pterygium is an extremely common disease of the surface of the eye which, if not diagnosed and treated early enough, can cause impaired vision. Correct detection, lesion segmentation and severity grading are crucial for proper clinical management. This paper introduces a hybrid CNN–Transformer model called EyePterNet-X for automated analysis of pterygium in anterior segment eye images. To enhance the global contextual representation of the Swin Transformer and the local feature extraction ability of EfficientNetV2, a novel Cross-Attention Guided Feature Fusion Module (CAFFM) is introduced to the proposed model. A unified multi-task learning framework learns a shared feature representation for pterygium detection, lesion segmentation and severity grading. The proposed approach is experimentally evaluated on 5,120 clinically annotated images, which proves the efficacy of the proposed approach. Based on the test results, EyePterNet-X achieves the highest detection accuracy of 96.2% (AUC of 98.5%), highest Dice coefficient of 87.5% for lesion segmentation, and the highest accuracy of 86.0% for grading of existing CNN-, Transformer-, and conventional hybrid-based methods. The findings show the validity, efficiency and reliability of the EyePterNet-X along with its ability to make a computer-aided pterygium diagnosis and large-scale ophthalmic screening.