Title: Robust Verification and Uncertainty Estimation with Deep Neural Networks.
Retinopathy is one of the major causes of blindness in the working-age population across the world. Since early diagnosis is both critical for successful treatment as well as a time-consuming and expensive process, there has been a surge in interest in developing ML aided methods for large scale screening. However, as is common in medical settings, the datasets available are highly imbalanced and noisy, necessitating the development of robust and interpretable methods.
Collaborating with Zeiss in this project we studied the questions
– is robustly training deep neural networks for diagnostics possible and scalable ?
– are the learned features meaningful for further interpretatino by medical experts?
– can we quantify the uncertainty and robustness of with the features in a meaningful way?
and were able to positively answer all three questions. The insights from this project are now being developed further and might see productive use in the future.