Advancing Stroke Lesion Segmentation With U-Net Variants: Classical, Transfer Learning, and MRI Sequence-Specific Customized Approaches

Abstract

Ischemic stroke, caused by restricted cerebral blood flow, is a leading cause of death and disability worldwide. Accurate segmentation of stroke lesions in MRI is essential for timely diagnosis but remains labor-intensive when done manually. This study presents a comparative evaluation of three U-Net variants for automated ischemic stroke lesion segmentation using the ISLES 2022 dataset: (1) Classical U-Net with a standard encoder-decoder structure, (2) Transfer Learning-Enhanced U-Net with MobileNetV2 encoder pre-trained on ImageNet, and (3) A novel MRI Sequence specific Customized U-Net that employs separate modality-specific encoders for DWI,ADC and FLAIR sequences followed by fused decoding. All models were trained and evaluated using Dice Score and Dice Loss metrics. The proposed customized U-Net outperformed the other two models in a single train-validation setup, achieving a Training Dice Score of 0.8680 and a validation Dice Score of 0.8409. The architecture demonstrates robust, efficient, and accurate segmentation, addressing class imbalance and small lesion challenges. These findings highlight the potential of modality-specific architectures to enhance clinical workflows and support automated stroke diagnosis.