A0150
Title: Statistical learning on AI-aided myopia screening
Authors: Catherine Liu - The Hong Kong Polytechnic University (Hong Kong) [presenting]
Abstract: Myopia has been a major public health concern in East Asia, especially in China, due to its high prevalence and potential vision-threatening complications. However, manual myopia screening is impractical at the population level due to the unaffordable healthcare costs. Luckily, AI-aided myopia screening makes it possible to do population myopia screening with operational expedience. We focus on ocular image-based AI-aided myopia screening: i) The first ophthalmic AI model that jointly considers axial length (AL) and myopia status and enhances the predictive capabilities; ii) the first ophthalmic AI model that handles the pair of eye-input images in AL and spherical equivalent (SE) predictions; and iii) the first vision transformer (ViT)-based ophthalmic model that diagnoses high-myopia and measures AL for both eyes. Theoretical interpretations of the AI models are contributed under multi-task learning, including: a) conditional dependence between responses can be modeled and interpretable in a convolutional neural network; b) interpretations of the fine-tuning procedure of ViT, the source of conditional dependence, and how the dependency benefits the ViT fine-tuning. This may provide evidence that statistics can harness AI and consequently benefit society in the data science and AI era.
