Title: A stroke lesion segmentation method based on volume-balanced data partitioning and dual-branch ensemble network

Authors: Siyu Zhao; Baoqiang Li

Addresses: Anshan Normal University, No. 43, Ping'an Street, Tiedong District, Anshan City, Liaoning Province, China ' Anshan Central Hospital, No. 77, South Zhonghua Road, Tiandong District, Anshan City, Liaoning Province, China

Abstract: Accurate segmentation of ischemic stroke lesions from MRI is crucial for clinical decision-making, including subtype classification and prognosis assessment. However, the heterogeneous size and appearance of lesions in T1-weighted MRI, along with class imbalance, pose significant challenges. In this study, we propose a dual-branch ensemble framework integrating nnU-Net and nnResU-Net to leverage their complementary strengths in global representation and local detail preservation. Furthermore, we introduce a volume-balanced cross-validation strategy to ensure consistent distribution of lesion sizes across training folds, addressing the imbalance problem at the data level. Experiments on the publicly available ATLAS R2.0 dataset demonstrate the superiority of our method. Our ensemble approach achieves a Dice score of 0.6601, volume difference (VD) of 9188 mm3, lesion-wise F1-score (L-F1) of 0.5349, and a simple lesion count (SLC) error of 4.6735 across five-fold cross-validation. These results outperform state-of-the-art baselines, including U-Net, TransUNet, and SwinUNETR. Qualitative visualisation further confirms that our model produces lesion segmentation results most closely aligned with expert annotations. To further enhance clinical applicability, the framework can be deployed in edge-computing environments, enabling low-latency and resource-efficient lesion segmentation close to the point of care.

Keywords: ischemic stroke; magnetic resonance imaging; deep learning; medical image analysis; segmentation.

DOI: 10.1504/IJBIDM.2026.155243

International Journal of Business Intelligence and Data Mining, 2026 Vol.28 No.4/5/6, pp.304 - 316

Received: 08 May 2025
Accepted: 13 Jan 2026
Published online: 29 Jul 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article