Development of an artificial intelligence model to identify duodenal polyps in patients with familial adenomatous polyposis

开发一种人工智能模型来识别家族性腺瘤性息肉病患者的十二指肠息肉

阅读:3

Abstract

BACKGROUND AND AIMS: Precancerous duodenal polyps can be subtle in familial adenomatous polyposis (FAP). Artificial intelligence models can detect gastrointestinal (GI) tract pathology; thus, we aimed to develop a model to identify duodenal polyps in FAP patients. METHODS: Images of duodenal polyps from upper GI endoscopic surveillance in FAP patients were obtained. Polyps were manually annotated (Label Studio; HumanSignal, San Francisco, Calif, USA), and the nnU-Net framework (Applied Computer Vision Lab of Helmholtz Imaging and the Division of Medical Image Computing at the German Cancer Research Center, Heidelberg, Germany) was used to automate polyp segmentation. Images were randomly divided into training, validation, and test sets (80%, 10%, and 10%, respectively). Primary performance metric was Dice coefficient (score 0-1). Manual counting was also used to compare the model's ability to identify polyps. RESULTS: Mean polyps per image by manual count was 4.2 (standard deviation [SD] = 4.8) and by prediction model 4.0 (SD = 4.7). Sixty of 87 images (69.0%) had 0 missed polyps, 74 of 87 (85.1%) missed a maximum of 1 polyp, and 83 of 87 (95.4%) missed a maximum of 4. Mean missed polyps per image was 0.9 (SD = 2.3), and the model identified 287 of 365 (78.6%) polyps. Forty of 87 (46.0%) had falsely identified polyps (mean 0.8; SD = 1.0). Missed polyps were smaller than identified polyps (mean 73.3 [SD = 45.3] versus 233.5 [SD = 274.5] pixel diameter, respectively). There was a complete match (no missed or false positive) in 33 of 87 images (37.9%). Dice coefficient was 0.73. CONCLUSIONS: A model to identify duodenal polyps in FAP was successfully created. Although the Dice coefficient is modest compared with that of colon polyp models, duodenal anatomy creates a challenging background for human and computer detection. Rate of polyp detection, likely a superior marker of goal achievement, was >75%, with a low false polyp rate (mean <1/image). This prototype model is the first step toward a refined algorithm to assist in identification of duodenal polyps with a need for larger prospective studies.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。