AMSSE-SwinTrans: A harmonized multimodal hybrid framework with explainable AI for ischemic stroke lesion segmentation.
Authors
Affiliations (2)
Affiliations (2)
- School of Computing Sciences, Hindustan Institute of Technology and Science, Chennai, India.
- Operative Research Unit of Neurology, Fondazione Policlinico Universitario Campus Bio-Medico, Rome, Italy.
Abstract
BackgroundAccurate ischemic stroke lesion segmentation from magnetic resonance imaging (MRI) is important but challenging due to variability in lesion morphology. Traditional convolutional neural networks (CNNs) effectively capture local spatial features but often struggle to capture long-range contextual dependencies needed for irregular lesions. Transformer-based architectures are good at capturing global context; however, they are computationally expensive and difficult to deploy in resource-constrained environments. The black-box nature of deep learning models limits their clinical adoption owing to lack of transparency.ObjectiveThis study introduces AMSSE-SwinTrans, a lightweight hybrid CNN-Transformer framework to make stroke lesion segmentation accurate, efficient and interpretable across different MRI modalities.MethodsThis study proposes AMSSE-SwinTrans, a lightweight hybrid CNN-Transformer framework that combines adaptive multi-scale squeeze-and-excitation (AMSSE) module with simplified transformer blocks employing multi-head self-attention, along with explainable AI (XAI) for clinical interpretation. The model was evaluated on a single-channel harmonized multimodal two-dimensional MRI Slices from ATLAS v2.0 and ISLES 2022 dataset.ResultsThe model achieved an overall Dice of 0.8629, lesion-specific Dice of 0.7535. An ablation study demonstrated a significant performance improvement when each component was added. The model comparisons were all significant with Wilcoxon signed-rank test yielding p-value (<i>p</i> < 0.001) across all comparisons. With only 27.67 M parameters, 17.67 GFLOPs, and 9.22 ms inference time per slice, the architecture is well-suited for deployment in resource-constrained clinical settings. The interpretability of the model is supported by Attention Rollout and Grad-CAM++ visualizations.ConclusionAMSSE-SwinTrans establishes itself as a resource-efficient and interpretable framework for multimodal stroke lesion segmentation.