VRXU-net outperforms existing models in ischemic stroke lesion segmentation

Sayeh Amir Mousavi Mobarakeh's VRXU-net tackles the challenge of detecting and segmenting brain ischemic stroke lesions in T1W MRI scans. These lesions vary widely in shape and size and often resemble healthy tissue, making manual segmentation difficult. The model first applies a modified VGG neural network to classify whether a 2D slice contains a lesion. Then, a U-shaped segmentation architecture with residual blocks (ResUnet) delineates lesion boundaries. Critically, VRXU-net processes slices separately from axial, sagittal, and coronal planes and aggregates the results for a 3D output. A high-performance classifier is placed before the segmentation network to skip non-lesion slices, reducing computational overhead and improving speed. This sequential design also lets segmentation feedback reduce false positives.

Trained on the Anatomical Tracings of Lesions After Stroke (ATLAS) dataset, VRXU-net beats current state-of-the-art models on both accuracy and Dice similarity coefficient—a metric for segmentation overlap. The multi-plane approach reduces ambiguity from grayscale similarity, and the residual blocks help capture fine-grained boundaries. The model's ability to filter out non-lesion slices makes it practical for clinical workflows where speed is critical. While the paper focuses on T1W MRI, the architecture could be adapted to other modalities. If validated in real-world clinical settings, VRXU-net could give radiologists a robust second opinion for stroke lesion assessment, aiding treatment decisions and surgical planning.