Abstract
Medical image segmentation task can provide the lesion object semantic information, but ignores edge texture details from the lesion region. Conversely, the medical image reconstruction task furnishes the object detailed information to facilitate the semantic segmentation through self-supervised learning. The two tasks are supplementary to each other. Therefore, we propose a multi-interactive feature embedding learning for medical image segmentation. In the medical image reconstruction task, we aim to generate the detailed feature representations containing rich textures, edges, and structures, thus bridging the low-level details lost from segmentation features. In particular, we propose an adaptive feature modulation module to efficiently aggregate foreground and background features to obtain a comprehensive feature representation. In the medical segmentation task, we propose a bi-directional fusion module fusing all important complementary information between the two tasks. Besides, we introduce a multi-branch visual mamba to capture structural information at different scales, thus enhancing model adaptation to different lesion regions. Extensive experiments on four datasets demonstrate the effectiveness of our framework.