Predicting potential diseases from retinal images
Fundus retinal imaging has been recognized as an essential tool for diagnosing various diseases not only of eyes but also of the circulatory system. Automating this diagnosis process can unload medical doctors' burden and may help reduce errors. Poplin et al. [Poplin et al., "Prediction of cardiovascular risk factors from fundus retinal photographs via deep learning," Nature biomedical engineering, 2019] has shown systematic analysis of the capability of fundus retinal images to predict various cardiovascular risk factors.
On top of this finding, we set three research questions: (i) Can we further improve prediction performance? (ii) Can we design a new metric that can be computed from fundus retinal images by integrating risk factors, and (iii) Can we transfer the trained model to predict other clinical entities? We aim at answering these research questions simultaneously by using deep neural network-related technologies. Our project focuses on how much we can push up the performance of such predictions.
Our project is intended to build a system that supports diagnosis by medical doctors by providing predicted risk factors for screening. Such systems can decrease medical doctors' burden with possibly reducing errors in diagnosis. This may contribute to building a fully automated system for quick screening of cardiovascular diseases by just looking into a fundus retinal image capturing device. We also try to produce a new integrated metric (or a risk factor) that can be directly extracted from a fundus retinal image. We consider that this will be a more robust and reliable metric, given that all existing risk factors have correlation to each other to some extent, as the training of this metric can use all observed risk factors associated with each sample.
At this moment, our project duration is two years.