Abstract:
Medical image biomarkers aid diagnosis and treatment. Supervised deep learning can accurately segment pathological regions, but it requires a priori definitions, large-scale annotations, and a representative patient cohort in the training set.
Contrary to pathology detection, anomaly detection can train on healthy samples without annotation. Biomarkers can be found in anomalies. Knowing normal anatomy helps detect anomalies. Bayesian deep learning can exploit this property by assuming that epistemic uncertainties correlate with anatomical deviations from a normal training set.
Bayesian U-Nets are trained on well-defined healthy environments using weak labels of healthy anatomy from existing methods. Monte Carlo dropout estimates model epistemic uncertainty at test time. A novel post-processing technique exploits these estimates and converts their layered appearance to smooth blob-shaped anomaly segmentations.
Using weak retinal layer labels, we validated this approach in retinal OCT images. Our method had a Dice index of 0.789 in an independent anomaly test set of AMD cases.
Our segmentations accurately separated healthy and diseased cases of late wet AMD, dry geographic atrophy (GA), diabetic macular edema (DME), and retinal vein occlusion (RVO). We also found that our approach can detect cut edge artifacts in normal scans.
Note: Please discuss with our team before submitting this abstract to the college. This Abstract or Synopsis varies based on student project requirements.
Did you like this final year project?
To download this project Code with thesis report and project training... Click Here
