PROJECT TITLE :

Exploiting Epistemic Uncertainty of Anatomy Segmentation for Anomaly Detection in Retinal OCT

ABSTRACT:

The detection of useful biomarkers in medical imaging aids in the diagnosis and treatment of patients. An accurate segmentation of diseased areas using supervised deep learning relies on prior definitions of these regions, wide-scale annotations, a representative patient cohort in the training set, and a large number of patients with similar characteristics. Although pathology is defined in terms of disease, anomaly detection allows for training on healthy data without any annotation. Biomarker discovery can then take place in these areas. This knowledge provides implicit clues for spotting abnormalities in the body's structure. Bayesian deep learning is proposed to take advantage of this trait, assuming that epistemic uncertainty are linked to anatomical deviations from a normal training set. U-Nets trained on a well-defined healthy environment are taught using weak anatomical annotations supplied by previous approaches. We use Monte Carlo dropout to capture our model's epistemic uncertainty estimates at testing time. After that, a new post-processing method is used to take advantage of these estimations and convert their layered look into smooth blob-shaped segments of the anomalies. Using OCT images of retinal layers with weak labelling, we were able to demonstrate the effectiveness of this method. Using an independent anomaly test set of AMD cases, our technique achieved a Dice index of 0.789. Finally, we found that our technique can also detect additional abnormalities in normal scans, such as cut edge artefacts.