Uncertainty for prediction quality assessment in clinical settings

Deep learning has the potential to increase efficiency of many routine tasks in medical image analysis. For instance, segmentation, the detection of the boundaries of specific target structures like organs or tumors, is a tedious and time-consuming chore for clinicians. Robust deep learning models could assist medical personnel for this type of task. However, automatic segmentations using machine learning and in particular deep learning could lead to medical errors due to so-called silent failure. If a model fails to produce a reasonable output segmentation it doesn’t give a warning to the user. To avoid this issue, verification of each single case from experts is required. Associating each prediction with an uncertainty value can indicate which predictions need more attention from the medical specialists. Hence, exploring uncertainty estimation methods that correlate with segmentation quality (e.g., Dice score) would greatly benefit the medical field and the neuroscience community. This project aims to study factors impacting uncertainty, hence will lead to increase insight on the uncertainty mechanisms associated with neural network. The ultimate goal is to find concrete methods to evaluate uncertainty to identify particularly challenging cases likely to result in low-quality segmentations.

Faculty Supervisor:

Julien Cohen-Adad

Student:

Partner:

Harvard Medical School

Discipline:

Computer science

Sector:

Education

University:

Polytechnique Montréal

Program: