Post-Stroke Lesion Segmentation using 2D and 3D Dense U-net

Strokes are brain attacks that cause death. Survivors suffer from damage to their brain (known as lesions). Detecting and segmenting these lesions using imaging techniques is crucial for medical diagnosis and treatment. However, manual post-stroke lesion segmentation is challenging and not yet fully automated. This research aims to assess the effectiveness of the Dimensional Dense U-Net architecture in this task. The architecture has shown promising results in other segmentation tasks, and we aim to evaluate its performance on the Anatomical Tracing of Lesion after Stroke (ATLAS) dataset. Initially, we employed a 3D version of the model and later switched to a 2D version when the 3D results were unsatisfactory. Among numerous trials, the 3D dense architecture achieved its highest dice score of 13% on patients with a single lesion in the left hemisphere, using 110 images. The 2D dense architecture achieved the highest dice score among all using 1000 random images. Each set was split into 70% for training, 20% for validation and 10% for testing. The training dice score was 87%, with a validation dice score of 9.8%.