Ischemic Stroke Detection, Segmentation, and Volume Estimation from Multi-sequence MRI with Missing Sequences
Synopsis
This work presents ISDS-MRI, a unified framework that uses graph neural networks and sequence-specific feature modeling to detect ischemic stroke, segment lesions, and estimate lesion volume from multi-sequence MRI while accommodating incomplete sequence combinations; evaluated across multiple public MRI datasets and a newly curated BGD-MRIS dataset of 532 scans from three hospitals in Bangladesh, it achieves a Dice score of 0.725, an AUC of 0.962, and a lesion volume estimation relative error of 8.4%, outperforming comparison methods by 3.2% in Dice, 2.6% in detection, and reducing volume relative error by 1.9%.
Interpretation
Introduces ISDS-MRI, a unified framework that integrates ischemic stroke detection, lesion segmentation, and lesion volume estimation, using graph neural networks and sequence-specific feature modeling. Whereas prior work often treats detection, segmentation, or volume estimation separately, this work brings the three tasks into one framework and explicitly models features of different MRI sequences. The paper presents this as a unified framework and reports quantitative results for all three tasks across datasets, including Dice 0.725, AUC 0.962, and volume relative error 8.4%.
The framework accommodates incomplete combinations of MRI sequences, performing stroke analysis even when some sequences are missing. Most multi-sequence MRI methods assume a complete set of input sequences, whereas this work treats sequence absence as a design condition. The abstract explicitly states the framework accommodates incomplete combinations of MRI sequences, and the multi-dataset evaluation reflects its robustness.
Constructs and introduces BGD-MRIS, a dataset of 532 MRI scans from three hospitals in Bangladesh, as a multi-center cohort from a resource-constrained setting. It adds a multi-center test bed from a resource-constrained region, spanning three hospitals and heterogeneous imaging protocols, for stroke AI evaluation. The paper reports the dataset size of 532 scans, its origin from three hospitals, and its use for evaluating stroke AI across heterogeneous clinical imaging protocols.
Evaluated on multiple public MRI datasets and BGD-MRIS, ISDS-MRI outperforms comparison methods on segmentation, detection, and volume estimation. Relative to comparison methods, Dice improves by 3.2%, detection performance by 2.6%, and volume estimation relative error decreases by 1.9%. The paper reports these comparison figures and highlights the results as reflecting the framework's robustness and clinical potential.
Perspective
The work targets ischemic stroke analysis from multi-sequence MRI and is meant for clinical and research settings where sequence combinations may be incomplete and imaging protocols heterogeneous, with particular attention to resource-constrained environments; its value lies in offering a unified, cross-dataset-evaluable framework for detection, segmentation, and volume estimation, and in adding BGD-MRIS as a multi-center test bed for resource-constrained regions.
The currently visible text is an abstract-level overview; it does not include network architecture details, training and validation splits, how sequence missingness was specifically simulated, statistical tests and confidence intervals, or the annotation process and case composition of BGD-MRIS. These details affect judgments of robustness and reproducibility and warrant consulting the full-text figures and supplementary materials. In addition, prospective validation of the framework in real clinical workflows and its performance across different scanners and populations remain to be observed.