Simi Meledathu Sasidharan: A Deep Residual U-Net Approach for Robust Stroke Lesion Segmentation in Diffusion-Weighted MRI

<p dir="ltr">Background and Aims: Precise segmentation of stroke lesions from diffusion-weighted imaging (DWI) is essential for quantifying ischemic damage and guiding post-stroke therapeutic interventions. This study presents the development and evaluation of a 3D Residual U-Net model for automated stroke lesion segmentation, focusing on achieving high accuracy and robust generalizability across heterogeneous patient data.</p><p dir="ltr">Methods: The model was trained using a dataset of 105 acute stroke DWI scans. A fivefold cross-validation strategy was applied to 80 training subjects to assess intra-dataset consistency.</p><p dir="ltr">Patch-based input sampling focused on lesion-containing regions to optimize learning efficiency. The model architecture incorporated deep residual connections and a strengthened bottleneck to enhance feature extraction. Evaluation metrics included the Dice similarity coefficient, Intersection over Union (IoU), precision, recall, F1-score, and area under the ROC curve (AUC). Generalization was tested on an independent held-out set of 25 subjects.</p><p dir="ltr">Results: The model demonstrated consistent cross-validation performance with Dice scores of 0.84, 0.82, 0.79, 0.84, and 0.82 across the five folds, indicating stable learning. On the held-out test set, the model achieved a Dice score of 0.8240, IoU of 0.70, precision of 0.85, recall of 0.798, F1-score of 0.82, and AUC of 0.89. These results confirm the model’s robust segmentation performance and ability to balance sensitivity and specificity.</p><p dir="ltr">Conclusions: The proposed 3D Residual U-Net achieves accurate and generalizable stroke lesion segmentation from DWI. Its strong performance supports its potential integration into automated neuroimaging pipelines for stroke diagnosis and clinical decision-making.</p>