Adaptive Group Fusion Module for Improving Feature Fusion of Multi-modality Lumbar Spine MRI

Guangtao Zhang,Han Yan,Zhuliang Yu
DOI: https://doi.org/10.1109/iaecst54258.2021.9695606
2021-12-10
Abstract:Automatic segmentation of lumbar spine tissue is essential in analyzing vertebral and intervertebral disc disorders. Recently, inspired by the success of deep learning in medical image segmentation, various deep learning methods have been applied in this task and have achieved remarkable progress. However, most deep learning methods rarely delve into multi-modal features of lumbar tissues. This paper introduces a novel feature fusion module called adaptive group fusion module (AGFM) to fuse features from different modalities for extracting potentially valuable cross-modality features. Furthermore, we design a novel multi-modality fusion network based on the UNet architecture to aggregate cross-modality features from low-level to high-level for lumbar spine segmentation. Extensive experiments on our multi-modality MR images of the lumbar spine demonstrate that our proposed AGFM outperforms the other feature fusion methods for multi-modality learning. Furthermore, we compared our network to other baseline multi-modality fusion segmentation networks. The experimental results show that our network has a significantly higher segmentation accuracy than the other baseline networks.
What problem does this paper attempt to address?