Establishment and test results of an artificial intelligence burn depth recognition model based on convolutional neural network
-
摘要: 目的 建立基于卷积神经网络的人工智能烧伤深度识别模型并测试其效果。 方法 在本诊断试验评价研究中,收集中南大学湘雅医院(下称笔者单位)2010年1月—2019年12月收治的符合入选标准的221例烧伤患者伤后48 h内创面照片484张,采用随机数字编号。采用图像查看软件圈出目标创面,由笔者单位烧伤整形科3名具有5年以上专科工作经验的主治医师判断烧伤深度,用不同颜色标记浅Ⅱ度、深Ⅱ度或Ⅲ度烧伤后,按224×224像素的尺寸切割得到完整大小的图像块5 637张。采用图片生成器将3种深度烧伤图像块均扩充至10 000张后,将每种烧伤深度图像块按7.0∶1.5∶1.5比例分为训练集、验证集和测试集。在Keras 2.2.4 Python 2.8.0版本下,采用卷积神经网络中的残差网络ResNet-50构建人工智能烧伤深度识别模型,输入训练集进行训练,利用验证集对模型进行调整、优化。利用测试集测试构建的模型识别各类烧伤深度的准确率,计算精确率、召回率及F1指数;通过降维工具tSNE将测试结果降维可视化生成二维tSNE云图,观察各类烧伤深度分布情况;根据模型对3种烧伤深度识别的敏感度及特异度,绘制出相应受试者工作特征(ROC)曲线,计算ROC曲线下面积。 结果 (1)经测试集测试,人工智能烧伤深度识别模型识别浅Ⅱ度、深Ⅱ度、Ⅲ度烧伤的精确率分别为84%(1 095/1 301)、81%(1 215/1 499)、82%(1 395/1 700),召回率分别为73%(1 095/1 500)、81%(1 215/1 500)、93%(1 395/1 500),F1指数分别为0.78、0.81、0.87。(2)tSNE云图显示,人工智能烧伤深度识别模型测试集测试结果中不同烧伤深度之间总体重叠较少,其中浅Ⅱ度与深Ⅱ度、深Ⅱ度与Ⅲ度烧伤之间重叠相对较多,而浅Ⅱ度与Ⅲ度烧伤之间重叠相对较少。(3)人工智能烧伤深度识别模型识别3种烧伤深度的ROC曲线下面积均≥0.94。 结论 采用ResNet-50网络建立的人工智能烧伤深度识别模型可较准确地识别烧伤患者早期创面照片中烧伤深度,特别是浅Ⅱ度与Ⅲ度烧伤,有望用于临床烧伤深度辅助诊断,提高诊断准确率。Abstract: Objective To establish an artificial intelligence burn depth recognition model based on convolutional neural network, and to test its effectiveness. Methods In this evaluation study on diagnostic test, 484 wound photos of 221 burn patients in Xiangya Hospital of Central South University (hereinafter referred to as the author′s unit) from January 2010 to December 2019 taken within 48 hours after injury which met the inclusion criteria were collected and numbered randomly. The target wounds were delineated by image viewing software, and the burn depth was judged by 3 attending doctors with more than 5-year professional experience in Department of Burns and Plastic Surgery of the author′s unit. After marking the superficial partial-thickness burn, deep partial-thickness burn, or full-thickness burn in different colors, the burn wounds were cut according to 224×224 pixels to obtain 5 637 complete wound images. The image data generator was used to expand images of each burn depth to 10 000 images, after which, images of each burn depth were divided into training set, verification set, and test set according to the ratio of 7.0∶1.5∶1.5. Under Keras 2.2.4 Python 2.8.0 version, the residual network ResNet-50 of convolutional neural network was used to establish the artificial intelligence burn depth recognition model. The training set was input for training, and the verification set was used to adjust and optimize the model. The judging accuracy rate of various burn depths by the established model was tested by the test set, and precision, recall, and F1_score were calculated. The test results were visualized to generate two-dimensional tSNE cloud chart through the dimensionality reduction tool tSNE, and the distribution of various burn depths was observed. According to the sensitivity and specificity of the model for the recognition of 3 kinds of burn depths, the corresponding receiver operator characteristics (ROC) curve was drawn, and the area under the ROC curve was calculated. Results (1) After the testing of the test set, the precisions of the artificial intelligence burn depth recognition model for the recognition of superficial partial-thickness burn, deep partial-thickness burn, or full-thickness burn were 84% (1 095/1 301), 81% (1 215/1 499) and 82% (1 395/1 700) respectively, the recall were 73% (1 095/1 500), 81% (1 215/1 500) and 93% (1 395/1 500) respectively, and the F1_scores were 0.78, 0.81, and 0.87 respectively. (2) tSNE cloud chart showed that there was small overlapping among different burn depths in the test results for the test set of artificial intelligence burn depth recognition model, among which the overlapping between superficial partial-thickness burn and deep partial-thickness burn and that between deep partial-thickness burn and full-thickness burn were relatively more, while the overlapping between superficial partial-thickness burn and full-thickness burn was relatively less. (3) The area under the ROC curve for 3 kinds of burn depths recognized by the artificial intelligence burn depth recognition model was ≥0.94. Conclusions The artificial intelligence burn depth recognition model established by ResNet-50 network can rather accurately identify the burn depth in the early wound photos of burn patients, especially superficial partial-thickness burn and full-thickness burn. It is expected to be used clinically to assist the diagnosis of burn depth and improve the diagnostic accuracy.
点击查看大图
计量
- 文章访问数: 162
- HTML全文浏览量: 35
- PDF下载量: 32
- 被引次数: 0