Optic Cup and Optic Disc Segmentation Based on improved TransUnet

Abstract

Glaucoma ranks as the second leading cause of blindness globally, surpassed only by cataracts. It inflicts irreversible harm to the optic nerve, and once vision is lost, restoration is unattainable for life[1]. Therefore, early detection of glaucoma is imperative.The cup-to-disc ratio serves as a primary diagnostic tool for identifying glaucoma. Generally, an excessively large cup-to-disc ratio in fundus photographs strongly suggests glaucoma. However, human error in diagnosing fundus photographs by clinicians is prevalent, leading to time-consuming, labor-intensive, expensive, and potentially inaccurate diagnoses during extensive screenings.To combat this challenge, an enhanced TransUnet-based method is proposed for optic cup and optic disc segmentation, aiding clinicians in large-scale glaucoma screenings. The model incorporates the Focus structure and CBAM structure. The Focus structure addresses information loss during single downsampling of images, preserving more data without changing image dimensions. Meanwhile, the CBAM structure integrates channel and spatial attention, enhancing the model's ability to extract features.On the ORIGA-650 dataset, this enhanced method achieved a mean Dice coefficient of 0.984 for the disc and 0.947 for the cup, along with a mean IOU value of 0.968 for the disc and 0.902 for the cup. Compared to alternative algorithms, the segmentation results exhibit superior accuracy, showcasing the efficacy of our proposed model.