Comparative Analysis of Deep Learning-Based Abdominal Multivisceral Segmentation

Abstract

The segmentation of multiple abdominal organs is essential for medical diagnosis and treatment of various abdominal conditions, such as surgical planning, image-guided interventions and diagnosis. The main challenges are the highly heterogeneous and complex anatomy, as well as the variability in size, shape and position of abdominal organs. And in recent years, deep learning techniques have been successfully applied to various medical image segmentation tasks. Therefore combining accurate and effective deep learning based segmentation methods is essential to obtain better clinical results. In this study, we present a comparison of three deep learning architectures for abdominal multi-organ segmentation, namely the Multiscale Attention Network (MA-Net), ResNet50-U-Net and U-Net++. We evaluated the performance of these three architectures on an abdominal MRI dataset consisting of different pathological and anatomical conditions. Our results show that MA-Net equipped with a multiscale attention mechanism outperforms ResNet50-U-Net and U-Net++ in terms of Dice coefficient, Jaccard index and Hausdorff distance. By effectively capturing and integrating multi-scale contextual information, MA-Net can better depict complex organ boundaries in the dataset. Therefore, the application of MA-Net or its variants to abdominal organ segmentation has the potential to significantly enhance clinical decision-making and patient care.