A Semantic Segmentation-Based Automatic Pterygium Assessment and Grading System

Qingbo Ji ¹ Wanyang Liu ² Qingfeng Ma ¹ Hui He ² Lin Zhang ² Lijun Qu ^2*

• ¹ Harbin Engineering University, Harbin, Heilongjiang Province, China
• ² Harbin Medical University, Harbin, Heilongjiang, China

Pterygium is a common eye disease, and accurately assessing its severity and implementing appropriate treatment plans can minimize patient suffering. To address the increasing number of pterygium patients and the relatively limited number of ophthalmologists, it is crucial to develop efficient AI-based diagnostic methods. This paper proposes an automatic pterygium grading system that combines deep learning models with traditional image processing techniques. The system consists of two main components: a semantic segmentation module and a severity assessment module. In the semantic segmentation module, we improved the Transunet network for the specific task, enhancing segmentation performance. In the severity assessment module, we used an improved curve fitting method to calculate the depth of pterygium invasion into critical regions of the eye, serving as a basis for assessing pterygium severity.Our research is based on slit-lamp microscope images of the anterior segment provided by the Ophthalmology Department of the Second Affiliated Hospital of Harbin Medical University. These images were annotated at the pixel level for training the semantic segmentation model. The segmentation model achieved an average Dice coefficient of 94.89 on the test set (with a Dice coefficient of 90.41 for the pterygium class). In the overall evaluation phase, we used additional photos for validation. The system's grading accuracy and weighted F1 score reached 0.9360 and 0.9363, respectively, with the Kappa consistency coefficient between the system and professional physician assessments reaching up to 0.8908, demonstrating the reliability of this method.