Computer Vision News 38 Deep Learning for the Eyes labeled data, the whole network gets trained without freezing any layers. The downstream tasks include a variety of experiments on ocular disease classification (diabetic retinopathy, glaucoma), ocular disease prognosis (age-related macular degeneration) and oculomic prognosis (ischaemic stroke, myocardial infarction, heart failure, Parkinson’s disease). Yukun told me about the importance of not only including ocular tasks but also oculomic tasks. The first reason is quite technical: the goal was to verify the generalisability and adaptability of the foundation model. The second reason is more on the medical side - the eye is a window to the whole body's health condition and hence is an important organ in systematic disease understanding. Finally, an explainability concept that computes the relevancy for transformers is applied in order to highlight regions that contributed to the classification. The method first assigns local relevance based on the Deep Taylor Decomposition principle and then propagates these scores through the layers. Some example heat maps can be seen in the figure! If you want to reproduce/adapt/use RETFound, I have great news for you! The team released all codes (PyTorch and Keras) and the weight files which you can find in the Code availability section of their paper! They are also currently working on creating an application template together with software engineers from Google to minimize the operation required in use! More about AI for Ophthalmology
RkJQdWJsaXNoZXIy NTc3NzU=