arXiv reaDer
Toward Generalizable Multiple Sclerosis Lesion Segmentation Models
Automating Multiple Sclerosis (MS) lesion segmentation would be of great benefit in initial diagnosis as well as monitoring disease progression. Deep learning based segmentation models perform well in many domains, but the state-of-the-art in MS lesion segmentation is still suboptimal. Complementary to previous MS lesion segmentation challenges which focused on optimizing the performance on a single evaluation dataset, this study aims to develop models that generalize across diverse evaluation datasets, mirroring real-world clinical scenarios that involve varied scanners, settings, and patient cohorts. To this end, we used all high-quality publicly-available MS lesion segmentation datasets on which we systematically trained a state-of-the-art UNet++ architecture. The resulting models demonstrate consistent performance across the remaining test datasets (are generalizable), with larger and more heterogeneous datasets leading to better models. To the best of our knowledge, this represents the most comprehensive cross-dataset evaluation of MS lesion segmentation models to date using publicly available datasets. Additionally, explicitly enhancing dataset size by merging datasets improved model performance. Specifically, a model trained on the combined MSSEG2016-train, ISBI2015, and 3D-MR-MS datasets surpasses the winner of the MICCAI-2016 competition. Moreover, we demonstrate that the generalizability of our models also relies on our original use of quantile normalization on MRI intensities.
updated: Fri Oct 25 2024 15:21:54 GMT+0000 (UTC)
published: Fri Oct 25 2024 15:21:54 GMT+0000 (UTC)
参考文献 (このサイトで利用可能なもの) / References (only if available on this site)
被参照文献 (このサイトで利用可能なものを新しい順に) / Citations (only if available on this site, in order of most recent)
Amazon.co.jpアソシエイト