Automatic segmentation of multiple organs on CT images by using deep learning approaches

Medical image segmentation is a critical task in various medical applications, including disease diagnosis, treatment planning, and monitoring of treatment outcomes. In this chapter, we present a comprehensive study on medical image segmentation using deep learning (DL) techniques. This chapter presents a comprehensive evaluation of DL models for medical image segmentation of abdominal organs. Specifically, four organ datasets, including kidney, spleen, pancreas, and liver, are analyzed and evaluated using various DL architectures. Additionally, a combined dataset comprising all four organs is used for analysis to assess the models' performance on a multiorgan segmentation task. For each case study, we first provide a detailed description of the dataset used, including its characteristics, imaging parameters, and labeling scheme. We then describe the preprocessing steps used to prepare the data for segmentation, including image registration, normalization, and augmentation. Next, we present the DL model architecture used for segmentation, which is customized for each dataset. The architecture includes a contracting path that captures high-level features from the input image, a bottleneck layer that summarizes the learned features, and an expanding path that reconstructs the segmented image. The results of each case study are presented, including the best-performing model and its corresponding architecture. In general, the 3D U-Net architecture achieved the highest performance on all datasets and the combined dataset. Furthermore, the evaluation metrics demonstrate the high accuracy of the DL models, indicating their potential for clinical use. Overall, this chapter provides a valuable evaluation of DL models for abdominal organ segmentation in medical imaging. The results highlight the potential of these models for improving clinical diagnosis and treatment planning. Additionally, the methodology presented in this chapter can serve as a useful guide for future studies in medical image segmentation.