南方医科大学学报

南方医科大学学报 ›› 2023, Vol. 43 ›› Issue (5): 815-824.doi: 10.12122/j.issn.1673-4254.2023.05.18

• • 上一篇    下一篇

肾小球超微结构分割算法：基于区域级对比学习的深度模型

林国钰，张桢泰，路艳蒙，耿 舰，周志涛，路利军，曹 蕾   

  1. 南方医科大学生物医学工程学院//广东省医学图像处理重点实验室//广东省医学成像与诊断技术工程实验室，中心实验室，基础医学院，广东 广州 510515；广州华银医学检验中心，广东 广州 510515
  • 出版日期:2023-05-20 发布日期:2023-06-12

A region-level contrastive learning-based deep model for glomerular ultrastructure segmentation on electron microscope images

LIN Guoyu, ZHANG Zhentai, LU Yanmeng, GENG Jian, ZHOU Zhitao, LU Lijun, CAO Lei   

  1. School of Biomedical Engineering//Guangdong Provincial Key Laboratory of Medical Image Processing//Guangdong Provincial Engineering Laboratory for Medical Imaging and Diagnostic Technology, Central Laboratory, School of Basic Medical Sciences, Southern Medical University, Guangzhou 510515, China; Guangzhou Huayin Medical Laboratory Center, Guangzhou 510515, China
  • Online:2023-05-20 Published:2023-06-12

RichHTML

15

PDF

131

摘要： 目的 为了提高深度模型对电子显微镜图像中肾小球超微结构的分割性能，提出了一种基于超微结构语义相似性的区域级自监督对比学习方法USRegCon。方法 USRegCon使用大量无标记数据对模型进行预训练，该预训练过程包括3个步骤：（1）模型对图像中超微结构信息进行编码和解码，并根据超微结构的语义相似性自适应地将图像划分为多个区域；（2）依据所划分的区域，使用区域池化操作提取出每个区域的一阶灰度区域表示和深度语义区域表示；（3）对于一阶灰度区域表示，构建了灰度损失函数，目标为最小化区域内的灰度差异和最大化区域间的灰度差异。对于深度语义区域表示，构建了语义损失函数，目标为最大化表示空间中正区域对的相似性和负区域对的差异性。这两个损失函数将联合对模型进行预训练。结果 基于私有数据集GlomEM，USRegCon在肾小球滤过屏障三层超微结构的分割任务中，对基底膜、内皮细胞和足细胞均获得了良好的分割结果，Dice系数分别为85.69±0.13%、74.59±0.13%和78.57±0.16%。该结果优于现有的多种图像级、像素级和区域级自监督对比学习方法，并逼近基于大规模标记数据集ImageNet的全监督预训练方法。结论 USRegCon促进模型从大量无标记数据中学习有益的区域表示，弥补了标记数据不足的缺陷，提升了模型对肾小球超微结构的识别和边缘分割能力。

关键词: 肾小球超微结构分割；电子显微镜；标记数据稀缺；自监督对比学习

Abstract: Objective We propose a novel region-level self-supervised contrastive learning method USRegCon (ultrastructural region contrast) based on the semantic similarity of ultrastructures to improve the performance of the model for glomerular ultrastructure segmentation on electron microscope images. Methods USRegCon used a large amount of unlabeled data for pre- training of the model in 3 steps: (1) The model encoded and decoded the ultrastructural information in the image and adaptively divided the image into multiple regions based on the semantic similarity of the ultrastructures; (2) Based on the divided regions, the first-order grayscale region representations and deep semantic region representations of each region were extracted by region pooling operation; (3) For the first-order grayscale region representations, a grayscale loss function was proposed to minimize the grayscale difference within regions and maximize the difference between regions. For deep semantic region representations, a semantic loss function was introduced to maximize the similarity of positive region pairs and the difference of negative region pairs in the representation space. These two loss functions were jointly used for pre-training of the model. Results In the segmentation task for 3 ultrastructures of the glomerular filtration barrier based on the private dataset GlomEM, USRegCon achieved promising segmentation results for basement membrane, endothelial cells, and podocytes, with Dice coefficients of (85.69 ± 0.13)% , (74.59 ± 0.13)% , and (78.57 ± 0.16)% , respectively, demonstrating a good performance of the model superior to many existing image-level, pixel-level, and region-level self-supervised contrastive learning methods and close to the fully- supervised pre-training method based on the large- scale labeled dataset ImageNet. Conclusion USRegCon facilitates the model to learn beneficial region representations from large amounts of unlabeled data to overcome the scarcity of labeled data and improves the deep model performance for glomerular ultrastructure recognition and boundary segmentation.

Key words: glomerular ultrastructure segmentation; electron microscopy; labeled data scarcity; self-supervised contrastive learning