An Enhanced UNet++ with InceptionNeXt Blocks and Feature-Scale Channel Attention for Ischemic Stroke Lesion Segmentation

Ischemic stroke lesion segmentation from Magnetic Resonance Imaging (MRI) remains a challenging task due to the small lesion size, irregular morphology, low contrast with surrounding tissue, and severe class imbalance. To address these challenges, this study proposes an enhanced UNet++ architecture that integrates Modified InceptionNeXt Blocks and Feature-Scale Channel Attention (FSCA) to improve multi-scale feature extraction and adaptive feature fusion. Hyperparameter tuning was performed by optimizing the number of initial filters, network depth, and loss function, followed by an ablation study to evaluate the contribution of each architectural component. The proposed model was implemented using the TensorFlow framework and evaluated on the ISLES 2022 dataset. To enhance lesion visibility, Diffusion-Weighted Imaging (DWI) was processed using Contrast Limited Adaptive Histogram Equalization (CLAHE) to generate enhanced DWI (eDWI). Three input configurations, namely DWI, DWI+ADC, and DWI+ADC+eDWI, were investigated through channel concatenation. Performance was assessed using Dice Score, Intersection over Union (IoU), Precision, and Recall. Experimental results demonstrate that the proposed UNet++ model with Modified InceptionNeXt Blocks and FSCA achieves the best performance using the DWI+ADC+eDWI configuration, obtaining a Dice Score of 0.8742, IoU of 0.8464, Precision of 0.9440, and Recall of 0.8906. Furthermore, the segmentation results exhibit high agreement with the ground truth, indicating that the proposed architecture effectively improves ischemic stroke lesion segmentation while maintaining computational efficiency.