Monte Carlo Concrete DropPath for Epistemic Uncertainty Estimation in Brain Tumor Segmentation

The paper of Natalia Khanzhina, Maxim Kashirin and Andrey Filchenkov “Monte Carlo Concrete DropPath for Epistemic Uncertainty Estimation in Brain Tumor Segmentation“ was presented at UNSURE workshop MICCAI 2021. In this work, the authors propose two novel model calibration frameworks for uncertainty estimation: MC ScheduledDropPath and MC Concrete DropPath. Particularly, MC ScheduledDropPath drops out paths in DNN cells during test-time, which has proven to improve the model calibration. At the same time, the MC Concrete DropPath method applies concrete relaxation for DropPath probability optimization, which was found to even better regularize and calibrate DNNs at scale. The scientiests investigate both methods on the problem of brain tumour segmentation and demonstrate a significant Dice score improvement and better calibration ability as compared to state-of-the-art baselines.