Atrous spatial pyramid pooling and multi-image data fusion for smooth muscle segmentation in upper gastrointestinal sphincters.

Gastrointestinal (GI) sphincters provide critical roles in regulating the transport of contents along the GI tract. Dysfunctions of GI sphincters are associated with a range of major digestive disorders. Despite their importance, the microstructures of GI sphincters are not well investigated. While micro-computed tomography (μ-CT) provides detailed 3D images, conventional segmentation methods rely on manual correction, which is both time-consuming and prone to human error. This study proposes a segmentation method using atrous spatial pyramid pooling (ASPP), which helps in capturing different effective fields of view from a given input feature map, thereof providing finer local and global information for a given pixel. Additionally, we explored the use of multi-species data fusion to make the model more generalized. The proposed segmentation network incorporating ASPP and multi-species data fusion improved the segmentation of sphincter muscle images. Specifically, it increased the dice score and Jaccard index by 3.7% and 5.8%, respectively, while reducing the variance compared to conventional methods.Clinical relevance- Techniques developed in this study will inform μ-CT segmentation of human upper GI sphincters for detailed structural analysis of muscular dysfunction.