Smartphone App to Detect Pathological Myopia Using Spatial Attention and Squeeze-Excitation Network as a Classifier and Segmentation Encoder

Abstract

Pathological myopia (PM) is a worldwide visual health concern that can cause irreversible vision impairment. It affects up to 20 crore population, causing social and economic burdens. Initial screening of PM using computer-aided diagnosis (CAD) can prevent loss of time and finances for intricate treatments later on. Current research works utilizes complex models that are too resource-intensive or lack explanations behind the categorizations. To emphasize the significance of artificial intelligence for the ophthalmic usage and address the limitations of the current studies, we have designed a mobile-compatible application for smartphone users to detect PM. For this purpose, we have developed a lightweight model, using the enhanced MobileNetV3 architecture integrated with spatial attention (SA) and squeeze-excitation (SE) modules to effectively capture lesion location and channel features. To demonstrate its robustness, the model is tested against three heterogeneous datasets namely PALM, RFMID, and ODIR reporting the area under curve (AUC) score of 0.9983, 0.95, and 0.94, respectively. In order to support PM categorization and demonstrate its correlation with the associated lesions, we have segmented different forms of PM lesion atrophy, which gave us intersection over union (IOU) scores of 0.96 and fscore of 0.97 using the same SA+SE inclusive MobileNetV3 as an encoder. This lesion segmentation can aid ophthalmologists in further analysis and treatment. The optimized and explainable model version is calibrated to develop the smartphone application, which can identify fundus image as PM or normal vision. This app is appropriate for ophthalmologists seeking second opinions or by rural general practitioners to refer PM cases to specialists.