Automatic detection of blood content in capsule endoscopy images based on a deep convolutional neural network
Automatic detection of blood content in capsule endoscopy images based on a deep convolutional neural network
复制标题
DOI：
10.1111/jgh.14941
复制
发表时间：
2020-07-01
影响因子：
4.1
通讯作者：
Tada, Tomohiro
中科院分区：
文献类型：
--
作者：
Aoki, Tomonori;Yamada, Atsuo;Tada, Tomohiro
背景与目的检测胃肠道血液含量是胶囊内镜的重要应用之一。可疑血液指示器（SBI）是一种常规工具，用于自动标记阅读系统中描述可能出血的图像。我们的目标是开发一个基于深度学习的系统来检测图像中的血液含量，并将其性能与SBI进行比较。方法我们使用27 847张CE图像（6503张来自29名患者的血液含量图像和21 344张来自12名患者的正常粘膜图像）训练深度卷积神经网络（CNN）系统。我们通过计算受试者工作特征曲线下面积（ROC-AUC）及其灵敏度、特异性和准确性评估了其性能，使用了10208张小肠图像（208张描绘血液含量的图像和10000张正常粘膜图像）的独立测试集。在单个图像分析中，使用相同的测试集将CNN的性能与SBI的性能进行比较。结果血药浓度测定的AUC为0.9998。在概率评分的截断值为0.5时，CNN的敏感性、特异性和准确性分别为96.63%、99.96%和99.89%，显著高于SBI（分别为76.92%、99.82%和99.35%）。经过训练的CNN需要250秒来评估10208张测试图像。结论我们开发并测试了基于CNN的CE图像血液含量检测系统。该系统具有优于SBI系统的潜力，并且需要对大型研究进行患者水平的分析。
Background and Aim Detecting blood content in the gastrointestinal tract is one of the crucial applications of capsule endoscopy (CE). The suspected blood indicator (SBI) is a conventional tool used to automatically tag images depicting possible bleeding in the reading system. We aim to develop a deep learning-based system to detect blood content in images and compare its performance with that of the SBI. Methods We trained a deep convolutional neural network (CNN) system, using 27 847 CE images (6503 images depicting blood content from 29 patients and 21 344 images of normal mucosa from 12 patients). We assessed its performance by calculating the area under the receiver operating characteristic curve (ROC-AUC) and its sensitivity, specificity, and accuracy, using an independent test set of 10 208 small-bowel images (208 images depicting blood content and 10 000 images of normal mucosa). The performance of the CNN was compared with that of the SBI, in individual image analysis, using the same test set. Results The AUC for the detection of blood content was 0.9998. The sensitivity, specificity, and accuracy of the CNN were 96.63%, 99.96%, and 99.89%, respectively, at a cut-off value of 0.5 for the probability score, which were significantly higher than those of the SBI (76.92%, 99.82%, and 99.35%, respectively). The trained CNN required 250 s to evaluate 10 208 test images. Conclusions We developed and tested the CNN-based detection system for blood content in CE images. This system has the potential to outperform the SBI system, and the patient-level analyses on larger studies are required.