Ischemic Stroke Detection, Segmentation, and Volume Estimation using Multi-sequence MRI Data with Missing Sequences

Abstract

Stroke remains one of the leading causes of disability and mortality worldwide, where timely and accurate diagnosis is critical for guiding treatment and improving patient outcomes. However, a global shortage of trained clinicians and radiologists continues to limit rapid and reliable interpretation of neuroimaging, particularly in resource-constrained settings. Artificial intelligence (AI) has emerged as a promising solution to this challenge by enabling efficient analysis of medical images. Here we present an Integrated Stroke Diagnosis System for MRI (ISDS-MRI), a unified framework that leverages graph neural networks and sequence-specific feature modeling for comprehensive ischemic stroke analysis. This framework is designed to detect ischemic stroke, segment lesions, and estimate lesion volume from multi-sequence MRI data, while accommodating incomplete combinations of MRI sequences. To ensure generalizability, we evaluate our approach across multiple publicly available MRI datasets and introduce a newly curated dataset, BGD-MRIS, comprising 532 MRI scans from three hospitals in Bangladesh. This newly curated dataset provides a multi-center MRI cohort from a resource-constrained setting, offering an additional test bed for evaluating stroke AI across heterogeneous clinical imaging protocols. Experimental results demonstrate that ISDS-MRI achieves a Dice score of 0.725 for lesion segmentation, a AUC of 0.962, and a lesion volume estimation relative error of 8.4%, outperforming comparison methods by 3.2% in Dice score and 2.6% in detection performance, while reducing volume estimation relative error by 1.9%. These results highlight the robustness and clinical potential of ISDS-MRI for scalable and comprehensive stroke diagnosis from MRI.