Lesion-Net: A Lesion-aware Hierarchical Transformer for Ischemic Stroke Segmentation in DWI
This preprint presents Lesion-Net, a two-dimensional hierarchical transformer for ischemic stroke lesion segmentation in diffusion-weighted MRI (DWI). The framework investigates whether preserving spatial detail and allocating greater encoder capacity to high-resolution stages can improve segmentation in patients with low lesion burden. Its architecture combines reduced early downsampling, greater early-stage depth, uniform channel widths, a multiscale decoder, and a feature-fusion segmentation head. Lesion-Net was evaluated on the ISLES 2022 and Johns Hopkins University Stroke (JHUS) datasets using DWI-only inputs and patient-level training, validation, and test partitions. It achieved overall Dice Similarity Coefficient scores of 79.82% on ISLES 2022 and 77.14% on JHUS. Patient-level burden analysis showed the highest small-burden Dice scores among the evaluated methods, reaching 81.80% and 80.32%, respectively. Architectural ablations examined the effects of spatial resolution, encoder-depth allocation, and channel configuration. The findings support adapting hierarchical encoder capacity to lesion characteristics. Lower performance relative to leading comparators in medium- and large-burden groups highlights a specialization trade-off and motivates further evaluation of lesion-level detection, cross-dataset generalization, and computational efficiency.
Authors
- Mohammad Alshurbaji (ORCID: https://orcid.org/0000-0003-0964-7786)
- Naoufel Werghi (ORCID: https://orcid.org/0000-0002-5542-448X)
- Taimur Hassan
- Maregu Assefa
- Mohamed L Seghier
- Kamal Taha
Institutions
- Abu Dhabi University (AE)
- Khalifa University of Science and Technology (AE)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24369221
- Primary Topic
- Medical Image Segmentation Techniques
- Type
- preprint