T1-Weighted Magnetic Resonance Imaging for Automatic Stroke Lesion Segmentation

Stroke is a major cerebrovascular disease, and as such, segmentation of lesions associated with stroke on MR images is of great interest to the medical imaging sector. Stroke lesion segmentation using magnetic resonance imaging (MRI) is a crucial step in determining the cause and severity of stroke symptoms. Our research proposes a deep learning architecture called 3D-Unet, which utilizes T1-weighted MRI images to differentiate stroke lesions. The 3D-Unet framework is developed as a 3D-compatible variation of the well-known U-Net structure. Using a skip-connected encoder-decoder network, high-resolution segmentation of volumetric data may be achieved. We used an ATLAS dataset annotated by radiologists to train our model on W1 MRI images with ground truth segmentation masks. Results demonstrate that the suggested 3D-Unet model provides state-of-the-art effectiveness in segmenting stroke lesions with a mean Dice coefficient of 0.81. Stroke lesion locations and sizes were successfully identified by the model, providing important information for medical practitioners. Promising results have been achieved by the suggested model in medical imaging segmentation tasks.