Brain Tumor Segmentation Leveraging Optimized U-net Models

Segmenting a brain tumor is a difficult project in the field of computer vision. It is of great importance in depicting the morphology of the tumor and hence assists clinicians to develop treatment plans. Machine learning, specifically deep learning techniques, has shown promising potential in image segmentation. By learning from large-scale datasets, the model could achieve useful information for identifying healthy and tumorous tissues. One of the most effective models is U-net. Since U-net is first proposed in 2015, it has been validated in multiple fields, especially for medical image processing. The core challenge of applying it to brain tumor segmentation is the complexity of the semantics it possesses. To seek the optimal model for the task, extensive ablation studies are conducted for searching the optimal hyperparameter combinations, measured by dice loss, IOU, and dice coefficient. Several modifications are implemented on the original built of U-net, including applying U-net++ and adding the attention module.