基于多特征拟合诊断胃褪色调肿瘤性病变的人工智能系统的构建和验证

Construction and validation of an artificial intelligence system based on multi‑feature integration for diagnosing gastric whitish neoplastic lesions

摘要:

目的 构建一个基于多特征拟合的人工智能诊断系统，用于白光内镜下诊断胃褪色调肿瘤性病变，并验证其诊断效能。方法 收集武汉大学人民医院和中国人民解放军总医院第七医学中心2012年11月至2021年7月的胃镜图像，选取267例患者的823张胃部褪色调病灶图像。通过文献检索选取5个与褪色调肿瘤性病变有关的白光内镜下特征：病灶位置、病灶边界是否清晰、病灶表面是否粗糙、病灶是否近圆形、病灶是否凹陷，把带有人工标注特征的图像输入机器学习算法中训练，选出最优模型作为多特征拟合诊断系统，系统确定各个病灶特征的诊断权重；用同样的数据集训练测试传统单一深度学习模型。对两个模型的诊断效能进行比较，并邀请8名不同级别的内镜医师进行人机比较。结果 多特征拟合诊断系统诊断胃褪色调肿瘤性病变的准确率、灵敏度、特异度分别是82.11%（101/123）、78.43%（40/51）和84.72%（61/72）。病灶特征按权重由高到低依次是病灶是否凹陷（权重0.71）、病灶位置（权重0.11）、病灶表面是否粗糙（权重0.08）、病灶边界是否清晰（权重0.06）和病灶是否近圆形（权重0.04）。多特征拟合诊断系统的诊断准确率显著高于非专家内镜医师平均诊断准确率（82.11%比74.31%，Z=-2.785，P=0.008），与专家内镜医师水平相当（82.11%比83.20%，Z=-0.696，P=0.700）。多特征拟合诊断系统与传统深度学习模型诊断准确率差异无统计学意义［82.11%（101/123）比82.93%（102/123），P=1.000］。结论 基于多特征拟合的人工智能诊断系统在白光下诊断胃褪色调肿瘤性病变具有较高的准确率。

Abstract:

Objective To construct and validate an artificial intelligence diagnostic system based on multi-feature integration for diagnosing gastric whitish neoplastic lesions under white-light endoscopy. Methods Gastroscopic images from Renmin Hospital of Wuhan University and the Seventh Medical Center of Chinese PLA General Hospital were collected from November 2012 to July 2021. A total of 823 images of gastric whitish lesions from 267 patients were finally selected. Five white-light endoscopic features associated with gastric whitish lesions were selected through a literature search, including lesion location, boundary clarity, surface texture, roundness, and depression status. Images with manually annotated features were used to train machine learning models, with the optimal model selected as the multi-feature fitting diagnostic system, which assigned diagnostic weights to each feature. A conventional deep learning model was trained with the same dataset. The diagnostic performance of the two models were compared, and eight endoscopists of varying expertise were invited to participate in human-machine comparisons. Results Accuracy, sensitivity, and specificity of the multi-feature fitting diagnostic system were 82.11% (101/123), 78.43% (40/51), and 84.72% (61/72), respectively. Feature weights in descending order were depression (0.71), lesion location (0.11), surface roughness (0.08), boundary clarity (0.06), and subcircular shape (0.04). The diagnostic accuracy of the system was significantly higher than that of non-expert endoscopists (82.11% VS 74.31%, Z=-2.785, P=0.008) and comparable to that of expert endoscopists (82.11% VS 83.20%, Z=-0.696, P=0.700). There was no significant difference in accuracy between the multi-feature fitting diagnostic system and the traditional deep learning model [82.11% (101/123) VS 82.93% (102/123), P=1.000]. Conclusion The feature-weighted artificial intelligence diagnostic system for gastric whitish neoplastic lesions demonstrates clinically relevant diagnostic accuracy under white-light endoscopy.