512层螺旋CT薄层扫描结合计算机人工神经网络对小肠病变的诊断价值研究  

The Value of 512 Slice Spiral CT Thin Layer Scanning Combined with Computer Artificial Neural Network in the Diagnosis of Small Intestinal Lesions

作　　者：廖忠剑[1] 文兴林[1] 刘艳平 邓星星 范存庚 雷剑[1] LIAO Zhongjian;WEN Xinglin;LIU Yanping(CT Room of Ganzhou People’s Hospital,Ganzhou City,Jiangxi Province 341000)

机构地区：[1]江西省赣州市人民医院CT室,341000

出　　处：《医学理论与实践》2024年第9期1458-1461,1470,共5页The Journal of Medical Theory and Practice

基　　金：赣州市卫生健康委员会市级科研计划项目(202005)。

摘　　要：目的:运用图像识别和图像描述技术,研究512层螺旋CT薄层扫描结合计算机人工神经网络在小肠病变诊断上的应用价值。方法:建立基于ResNet 101、Faster R-CNN、LSTM的小肠CT图像识别描述人工神经网络模型,采用ResNet 101和Faster R-CNN联合提取图像特征信息进行ROI融合建立图像编码,采用LSTM进行图像解码输出CT图像诊断报告。设计11个病变图像特征标签,选择我院2020年1月—2022年1月行512层螺旋CT薄层扫描病例材料1572套,材料覆盖正常图像及出血、炎性息肉等9种病变类型,按7∶3随机分为训练组和测试组,对模型进行训练并展开性能测试,评价模型输出的CT诊断文字报告质量、诊断准确率及病变部位提取正确率。结果:文字报告质量评分为(4.0325±0.684)分,出现错误评估项样本占比0.8%,仅语句结合约35%存在不合理现象;各病变类型诊断准确率均在97%左右,仅腺癌存在轻微的诊断为腺瘤性息肉现象,整体病变类型诊断正确率97.17%;病变部位正确识别提取率99.43%。结论:基于ResNet 101、Faster R-CNN、LSTM建立人工神经网络模型,进行小肠512层螺旋CT扫描图像的诊断识别,可准确定位图像病变部位,根据病变部位图像特征准确进行病变类型诊断,并就11个CT图像特征进行文字化描述,输出符合医师阅读理解的图像诊断报告,并发现部分医师人工阅片遗漏错误的地方,可为医师提供更为丰富、有效、准确的CT图像诊断信息,有利于医师作出正确的诊断。Objective:To study the application value of 512 slice spiral CT thin layer scanning combined with computer artificial neural network in the diagnosis of small intestinal lesions using image recognition and image description techniques.Methods:Establish an artificial neural network model for small intestine CT image recognition and description based on ResNet 101,Faster R-CNN,and LSTM.Use ResNet 101 and Faster R-CNN to jointly extract image feature information for ROI fusion and establish image encoding.Use LSTM to decode the image and output CT image diagnostic reports.Design 11 lesion image feature labels,select 1572 sets of case materials for 512 slice spiral CT thin layer scanning in our hospital from January 2020 to January 2022,covering 9 types of lesions such as normal images,bleeding,and inflammatory polyps.Randomly divide the model into a training group and a testing group at a ratio of 7∶3.Train the model and conduct performance tests to evaluate the quality of CT diagnostic text reports output by the model,diagnostic accuracy,and accuracy in extracting lesion sites.Results:The quality scores of the written report was 4.0325±0.684,with a sample proportion of 0.8%showing incorrect evaluation items.Only about 35%of the sentences were combined,resulting in unreasonable phenomena.The diagnostic accuracy of each lesion type is around 97%,with only mild adenomatous polyps being diagnosed in adenocarcinoma.The overall diagnostic accuracy of the lesion type is 97.17%.The correct recognition and extraction rate of lesion sites is 99.43%,which can accurately extract lesion site images for feature extraction.There are only slight errors in extracting lesion images in normal images.Conclusion:Based on ResNet 101,Faster R-CNN and LSTM,an artificial neural network model is established for the diagnosis and recognition of 512 slice spiral CT images of the small intestine.It can accurately locate the lesion site in the image,accurately diagnose the lesion type based on the image features of the lesion site,and provid

关 键 词：小肠病变 人工智能 薄层扫描CT 图像诊断描述 

分 类 号：R445[医药卫生—影像医学与核医学]