Back

Setting Standards to Promote Artificial Intelligence in Colon Mass Endoscopic Sampling

Zheng, Y.; Su, R.; Wang, W.; Meng, S.; Xiao, H.; Zhang, W.; Xu, H.; Bu, Y.; Zhong, Y.; Zhang, Y.; Qiu, H.; Qin, W.; Zhang, Y.; Xu, W.; Chen, H.; Zhang, C.; Wu, S.; Han, Z.; Zheng, X.; Zhu, H.; Wu, S.; Pan, W.; He, Y.; Hu, Y.

2019-10-08 oncology
10.1101/19008078 medRxiv
Show abstract

ObjectiveArtificial intelligence (AI) has undeniable values in detection, characterization, and monitoring of tumors during cancer imaging. However, major AI explorations in digestive endoscopy have not been systematically planned, and more important, most AI productions are based on Single-center Studies (ScSs). ScSs result in data scarcity, redundancy as well as island effects, which leads to some limitations in applying it on endoscopy. We investigate the disadvantages of picture processing which may effect the AI detection, and make improvements in AI detection and image recognition accuracy. DesignCurrent investigation aggregates a total of 2,500 gastroenteroscopy samples from various hospitals in multiple regions and carries out deep learning. ResultsIt is found that factors inconducive to AI recognition are common such as: (a) the gastrointestinal tract is not cleaned up completely; (b) shooting angle (from left to right and the top of polyp are unexposed clearly), shooting distance (too close or too far to shoot causes the lump to be unclear), shooting light (insufficient light source or overexposed light source in mass) and unstable shooting lead to poor quality of pictures. ConclusionWe set standards for a multicenter cooperation involving three-level medical institutions from the provincial, municipal and county to improve the recognition accuracy as well as the diagnosis and treatment efficiency meanwhile.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.