A project to use Mask R-CNN for brain tumor segmentation
Brain tumor segmentation is an essential step in the early diagnosis and treatment evaluation of brain tumor, but if done manually it is often time-consuming and the quality of the segmentation varies based on the operator’s experience. This leads to the requirement of a fully automatic method for brain tumor segmentation. In this work we explore the use of the Mask R-CNN framework to achieve this goal, with the idea to concentrate the segmentation on the detected region of interest, unlike conventional methods that perform the segmentation on the whole brain scan image, resulting in possibly reduced segmentation performance due to the disturbance of background regions. We tested different configurations to reach optimal performance. Our method is trained and evaluated on the BraTS 2020 dataset containing 369 brain tumor cases, and the results show promising performance.
Our code consists of two Jupyter notebooks, one for preprocessing the BraTs dataset and the other for training and evaluating the models using the Detectron2 library. We have developed functions for the Detectron2 library that take our generated dataset and plug it into the library to be compatible with the different detection models offered by the library. This means that our dataset can be easily used to train a variety of models such as Cascade R-CNN, rotated bounding boxes, PointRend, DeepLab, and ViTDet, by simply modifying the name of the model to be used. In addition, the preprocessing code converts the BraTs dataset from a semantic segmentation task to an instance segmentation task and generates an annotation file under the COCO format. Overall, our code is designed to be easily reusable, with a variety of models provided for training and evaluation.