EyeNomaly
A screening tool for retinal disease through unsupervised abnormality detection
Background
This project is a great opportunity for any student interested in deep unsupervised learning.
Eye2Gene is a deep-learning based AI algorithm for identification of Inherited Retinal Dystrophies (IRDs) using retinal imaging. In conjunction with the Eye2Gene classifier we have trained a Generative Adversarial Network (GAN) to generate images of retinal scans for patients with various IRDs. Prior works suggests that trained GANs can be adapted to perform anomaly detection In addition to identifying invalid data, anomaly detection algorithms trained on patients without any disease/dystrophies could be a good way of flagging up patients with a potential eye disorder for further investigation.
Goals:
- Identify out-of-distribution samples (i.e. data from a completely different kind of data to what the model was trained on) for the robust deployment of deep learning algorithms in the real world.
- Use anomaly detection to detect and reject out-of-distribution samples at run-time, and avoid giving the end-user predictions on data where the performance of the algorithm is unknown (and typically extremely poor).
Dataset
This project will use medical images from Moorfield inherited retinal diseases cohort [1]. The dataset contains images of fundus autofluorescence, infrared imaging, and OCT. These images belong to more than 100 types of IRDs which were identified using genetic diagnosis from an accredited genetic testing laboratory.
References
- https://pubmed.ncbi.nlm.nih.gov/32423767/
