Go to DBLP homepageGLMR-Net: Global-to-local mutually reinforcing network for pneumonia segmentation and classification
Abstract: The joint segmentation and classification tasks can simultaneously accomplish lesion localization and classification for pneumonia using CT images, assisting physicians in better diagnosis. However, a large number of existing methods can only accomplish the individual tasks mentioned above. In fact, both the segmentation and classification of pneumonia are focused on the lesion area, enabling the possibility of combining these tasks. Nonetheless, two difficulties persist that need to be primarily addressed: on one hand, the joint task requires more comprehensive feature extraction capability to adapt to the different tasks; on the other hand, the joint task involves more than simply combining two tasks, it necessitates an entire framework, thereby requiring effective mechanisms for interaction and mutual reinforcement. To address the aforementioned challenges, we propose a novel method called GLMR-Net, which can accomplish both segmentation and classification tasks and achieve mutual reinforcement between them. Specifically, GLMR-Net performs lesion feature information processing in a global-to-local manner and achieves feature-level and task-level mutual reinforcing. Our proposed GLMR-Net comprises three subnetworks: cross-feature extraction subnetwork, segmentation subnetwork, and classification subnetwork. Cross-feature extraction subnetwork realizes global feature extraction and feature-level interaction through Transformer-based backbone and our proposed cross-task differential feature fusion (CTDFF) module. Segmentation and classification subnetworks are used for local processing of features and generation of multi-scale lesion masks and lesion evidence masks used to construct task interaction loss (TI Loss) for task-level mutual reinforcement. Finally, we conduct extensive experiments and achieve superior performance compared to the existing state-of-the-art methods.
Loading