A Weakly Supervised Brain Tumor Segmentation Strategy Based on Multi-level Sub-category and Membership Matrix.

Abstract

Using a classification network to generate class activation mapping (CAM) is a mainstream method for weakly supervised semantic segmentation. However, for brain tumor images, CAM cannot fit the boundary of the tumor well. To improve the performance of brain tumor CAM, we propose a weakly supervised learning strategy based on a multi-level sub-category and membership matrix. Firstly, a multi-level sub-category strategy is used to intensively classify the data set. It allows the convolutional network to learn the in-depth characteristics of the input for enhancing CAM. Secondly, the idea of fuzzy clustering is introduced into model learning. The membership matrix is combined with CAM to construct the loss function. Exhaustive experiments on the brain tumor dataset BraTS2019 demonstrate that the proposed method can effectively improve the performance of CAM. Compared with the baseline method, our approach significantly improved by 17.1% using the common dice similarity coefficient evaluation approach, and compared with the recent study, our score also improved by almost 9%. The proposed methods train the network under image-level labels and help the convolutional network mine the target boundary information. They can help CAM fit the target border more accurately.