White Matter Multiple Sclerosis Lesion Segmentation Under Distributional Shifts

Abstract

In the rapidly evolving era of machine learning and deep learning, new algorithms are constantly emerging, each built upon existing research and pushing the boundaries in the field of medical imaging. However, one of the major challenges in the application of these algorithms is the distributional shifts that occur in real-world datasets. This research paper utilizes the expanded Shifts 2.0 dataset that was released for The Shift Challenge 2022. It presents how to enhance the UNET model’s robustness and uncertainty estimations in the segmentation of white matter lesions in Multiple Sclerosis patients, using only the FLAIR modality. This approach examines the impact of multiple hyperparameters on the results of the Shift 2.0 dataset. The suggested model yielded R-AUC scores of 1.12 and 1.60 on the Dev-out and Eval-out of the shift dataset, in contrast to the baseline UNET method which registered scores of 4.66 and 7.40 on those respective partitions. Moreover, the paper establishes that the performance of an ensemble of UNET models can be comparable to that of a transformer-based ensemble of UNETR models, offering promising implications for future research and applications.