Open Forum on Alzheimer’s and Small Vessel Disease Imaging
Presenter:
Lyndon Boone
Abstract:
A number of studies have shown that deep learning methods are capable of achieving near-human-level performance on neuroimaging segmentation tasks. With that said, most of the results quoted in the literature supporting this statement are in the context of test sets drawn from the same overarching dataset as the training data. Clinically-deployed models faced with out-of-distribution data (i.e. data that does not resemble the training set) may severely underperform relative to the standards set in the literature if they aren’t designed specifically with robustness to out-of-distribution data in mind. In this talk, Lyndon highlighted the sensitivity of modern CNN-based architectures to image corruptions, artifacts, and post-processing transforms commonly found in MRI. He then presented a methodology for benchmarking different architectures and models on the basis of robustness to out-of-distribution data, inspired by similar work in the computer vision literature.
Modality:
- MR
Organ:
- Brain