Gradual modality dropout for segmenting ischemic stroke lesions in an unseen center with missing modalities

Sofia Vargas Ibarra, Vincent Martin VIGNERON, Hichem MAAREF, Sonia Garcia-Salicetti, Andreia V Faria

Keywords: Stroke, segmentation, MRI modalities, lesion, missing modalities

Abstract: In clinical practice, imaging modalities may not always be available for every patient due to scheduling, cost, or patient-specific constraints. Additionally, multi-center imaging studies often face inconsistencies in protocols, machine settings, and artifacts, compromising data quality. We propose a 3D U-Net model for ischemic lesion segmentation using a novel training technique, gradual modality dropout, which progressively deactivates imaging modalities during training. This approach ensures robust performances when all modalities are present and improves segmentation accuracy in scenarios where one or more modalities are missing in unfamiliar contexts. The model demonstrates adaptability and reliability when trained on MRI scans of stroke patients across different phases (hyper-acute,sub-acute, acute, and post-treatment) and various hospital settings. Code available here: https://github.com/sofiavarib/Gradual-modality-dropout

Submission Number: 22