Current Issue Cover


摘 要
面对严重的医学影像分析缺口,深度学习的发展能够满足国内医疗行业的需求。本文回顾了2012年至2022年有关心室、心外膜、心包脂肪的图像处理的各项方法、衡量指标及其目前的研究现状,可大致分为传统的图像处理技术、基于图谱的方法(Atlas-based methods)、模型法(Model-based methods)、以及目前热门的采用机器学习和深度学习的方法。在深度学习兴起之前,传统的机器学习技术如模型法和图集法在心脏图像分割中的有良好表现。但是,它们通常需要大量的特征工程知识或先验知识才能获得令人满意的精度。相反基于深度学习的算法能从数据中自动发现复杂的特征以进行对象检测和分割。得益于先进的计算机硬件以及更多可用于训练的数据集,基于深度学习的分割算法已超越了以往的传统方法。最后结合分割技术的发展,讨论心脏分割的发展趋势。
Research status of cardiac image segmentation based on deep learning

Zeng Jia tao,Zhang He Ye,Liu Hua Feng(Sun Yat-sen University;Zhejiang University)

According to the World Health Organization, cardiovascular disease is the number one cause of death worldwide due to its high morbidity and severe sequelae. Facing the serious gap in medical image analysis, the development of deep learning can meet the needs of the domestic medical industry. Before the development of deep learning, most image processing was processed through thresholding. After the development of deep learning, people found that it is possible to guide the function to be close to reality by setting specific eigenvalues. In terms of its good learning ability and data-driven factors, compared with traditional methods, image processing based on deep learning is more excellent and more robust, such as based on deep residual network and generative confrontation network. This article analyzes the characteristics of representative methods, summarizes the resources and scale of cardiac images, compares the performance evaluation and application conclusions of different methods through commonly used evaluation indicators, and discusses the clinically applicable fields. The analysis papers in this article come from IEEE, SPIE, and China National Knowledge Network, with image processing and heart as search keywords. This review first classifies all methods according to different division areas, and further subdivides different methods in the same division area, and also provides information on related datasets. In this paper, the difference between image processing methods is evaluated by Dice coefficient and Hausdorff distance, and the performance of various methods is further quantitatively evaluated. In chamber segmentation, several approaches for right ventricle segmentation are reviewed. As far as the principle of the segmentation method is concerned, the method relying solely on the threshold cannot overcome the challenging problem of compartment segmentation unless it is combined with other methods, so it cannot be used as a single compartment segmentation technique. Therefore, in this part of the review, the existing research methods are mostly included in the second and third generation. Therefore, the following subsections review considering several categories. For each method, we first give a brief theoretical introduction to the method. Then, a detailed overview is given of each considered paper, its methodology, the datasets used, and the effectiveness of the evaluation segmentation process employed by the evaluation technique. Finally, the advantages and disadvantages of each method are summarized. In the chapter on epicardium and pericardium tissue, this article will briefly introduce the image processing techniques used to segment epicardium and pericardium tissue. We divide the papers analyzed in this paper into four groups according to the main methods: traditional image processing methods, atlas-based methods, machine learning, and deep learning. Traditional image processing methods include techniques such as thresholding, region growing, and active contouring. Finally, the capabilities of each algorithm are compared horizontally through the dice coefficient. For the segmentation method of the epicardium, it is easier to segment the epicardium in combination with outlining the pericardium than directly segmenting the epicardium. Epicardial and pericardial fat are unevenly distributed around the heart, resulting in large variability between sections and between patients on CT and MRI images. They are also non-uniform in shape and cannot be easily described in an algorithm. However, on CT and MRI images, the pericardium has a smooth, thin, almost oval outline. Methods such as active contours or ellipse fitting are naturally suitable for segmenting such shapes. Once the pericardium is divided, the epicardium is more easily divided into all the fatty tissue within the pericardium. The biggest challenge in segmenting the epicardium is how thin it is. When collecting CT scans for CAC scoring, the slice thickness is usually set at 2-3 mm. The pericardium is usually less than 2 mm thick, so due to partial volume averaging, the pericardium will often appear blurred or blurred on CT images, especially since the heart is a constantly moving organ. There are currently several methods for delineating the pericardium. Some methods are purely pericardial delineation methods, while others are part of a method to segment and quantify the epicardium. In the epicardium part, this article will mainly introduce the method of segmenting the epicardium by first segmenting the pericardium. Pericardial fat segmentation methods typically rely on traditional image processing methods, such as thresholding and region growing, using various preprogrammed heuristics to identify common structures and segment pericardial fat. Recent approaches employ atlas-based segmentation, but its clinical importance is relatively small. After introducing the current situation of segmentation, this article will introduce some real scenarios applied in clinical practice. Through these scenarios, we can see that cardiac image processing has a large number of clinical problems waiting to be solved. At the same time, this article will briefly introduce the market situation of image processing in the Chinese market, the combination of industry, education and research, and the main relevant policy trends. The third part will introduce the development of the main related industries, mainly referring to: the establishment of related imaging databases in China, the development of related imaging technologies in China, and the development of related hardware equipment in China. At the end, it is discussed that the development of cardiac image segmentation processing is increasingly inseparable from the development of deep learning. However, because deep learning itself is difficult to explain, models with more medical knowledge interpretation are needed, and the development of deep learning itself has great limitations. : More data sets with uniform standards and higher accuracy are needed.