Structural-Functional Transition in Glaucoma Assessment Edition2

This challenge is to predict the result of the visual field (VF) test using multi-modal ophthalmic images, including volume optical coherence tomography (OCT) scans centered at the fovea and color fundus photography (CFP). Unlike STAGE challenge in last year, in 2024, we want players to use multi-modal data to make predictions of the VF. OCT is now the most widely used imaging modality used in ophthalmology tests, which provides objective cross-sectional information of the structures in the fundus, facilitating the physician's observation of structural thickness changes, and is an important basis in the diagnosis of glaucoma. There is now much evidence to support the role of OCT imaging in the detection of glaucoma. CFP is the most common surface imaging mode of fundus structure, which is convenient and non-invasive. A VF test is a reference standard examination to assess visual function. It is a subjective examination that requires the subject to remain calm and focused and to cooperate with the physician. The monocular visual field examination takes approximately 15 minutes. It is the clinical standard to decide whether there is glaucomatous optic nerve damage. In contrast, a monocular volume OCT scan or CFP takes only about 3 seconds. Furthermore, there is a moderate to good correlation between retinal layer thickness calculated from OCT scans, or the cup-to-disc ratio, disc rim and retinal nerve fiber defect morphology observed in the CFP and central VF sensitivities or other markers of optic nerve function. Therefore, this challenge focuses on how to predict functional VF information using objective and easy-to-acquire structural OCT images and CFPs. Three tasks are proposed for this challenge: 1) mean deviation (MD) value prediction using CFP; 2) sensitivity map prediction using multi-modal data; 3) pattern deviation probability map prediction using multi-modal data. Our challenge will provide 400 volume OCT and CFP paired data and corresponding MD value, sensitivity map, and pattern deviation probability map labels of the VF test report. Of these, 200 multi-modal data and corresponding labels will be released to the teams for model training in the preliminary round. 100 multi-modal data will also be released in the preliminary round, and the evaluation platform will be opened for the teams to validate and tune their models based on the preliminary leaderboard The remaining 100 multi-modal data will be released in the final round for the evaluation of the model. From the technical point of view, this challenge is concerned with computer vision studies, among which, tasks 1 and 2 involve metric regression problems, and task 3 involves a classification problem. These studies are essential in computer-aided clinical diagnosis. From a biomedical perspective, this challenge is to seek the mapping relationship between the fundus structure and visual function, which is important for understanding the underlying causes of visual defects.