A Review of Semantic Medical Image Segmentation Based on Different Paradigms

Introduction

Medical images can intuitively reflect the anatomical structure and tissue function and extract large amounts of rich pathological information for medical image segmentation, classification, and disease detection (Johny et al., 2021). This assists doctors in treating diseases, surgery planning, and rehabilitation monitoring. Wang et al. (2018) proposed a cascaded U-Net network combined with a graphical model to segment the aorta, pulmonary artery, myocardium, left and right ventricles, and left and right atria of the heart and then performed similarity shape analysis for image comparison. In recent years, researchers have successfully implemented deep learning (DL) for brain (Zhuang et al., 2022), ear (Xu et al., 2019), liver (Fogarollo et al., 2023), spleen (Sharbatdaran et al., 2022), lung (Johny et al., 2021), kidney (Song et al., 2022), and multi-organ (N. Shen, et al., 2023) segmentation. DL has since been widely used in clinical applications. Due to the ability to extract and analyze biomedical information and apply image segmentation techniques to help doctors and researchers better understand diseases and human physiological conditions, medical image segmentation with DL has become a hot research topic.

Representative literature in the field of medical image segmentation includes the work of Ramesh et al. (2021) and Liu et al. (2021), who focused on classification of the fully convolutional network (FCN), U-Net, and mask R-CNN, but they only touched on the use of DL. Fu et al. (2020) combined image alignment and DL but derived only a few relevant combination techniques. Asgari et al. (2021) introduced FCNs, U-Net, and guided convolutional neural networks (CNNs) and classified them purely from the perspective of DL methods, with little elaboration. With the goal of addressing the shortcomings of previous research, in this study, we sought to provide a systematic exposition of DL-based medical segmentation methods. The study described in this paper contributes to the current literature in the following ways:

• 1)

  This study focused on DL-based medical segmentation methods. It systematically elaborates on basic concepts, basic methodological processes, and learning paradigms, with emphasis on relevant technologies used in medical segmentation and the latest technological improvements in the field.

• 2)

  Systematically discussed are the basic ideas of DL segmentation methods in the literature, a summary of the basic methods and main technologies of DL segmentation, and an analysis of the advantages and disadvantages of these technologies.

• 3)

  Common medical image segmentation datasets and evaluation metrics are comprehensively summarized. Then described are the data collection and annotation, model interpretability, multi-modal medical image segmentation, semi-supervised learning, unsupervised learning, and optimization of network structure.