https://www.ncbi.nlm.nih.gov/pubmed/29928897
Deep Learning Localizes and Identifies Polyps in Real Time With 96% Accuracy in Screening Colonoscopy.
Author information
- 1
- Department of Computer Science, University of California, Irvine, California; Institute for Genomics and Bioinformatics, University of California, Irvine, California.
- 2
- Department of Medicine, University of California, Irvine, California.
- 3
- Department of Medicine, University of California, Irvine, California; H.H. Chao Comprehensive Digestive Disease Center, University of California, Irvine, California.
- 4
- Department of Computer Science, University of California, Irvine, California; Institute for Genomics and Bioinformatics, University of California, Irvine, California; Center for Machine Learning and Intelligent Systems, University of California, Irvine, California. Electronic address: pfbaldi@uci.edu.
Abstract
BACKGROUND & AIMS:
The benefit of colonoscopy for colorectal cancer prevention depends on the adenoma detection rate (ADR). The ADR should reflect the adenoma prevalence rate, which is estimated to be higher than 50% in the screening-age population. However, the ADR by colonoscopists varies from 7% to 53%. It is estimated that every 1% increase in ADR lowers the risk of interval colorectal cancers by 3%-6%. New strategies are needed to increase the ADR during colonoscopy. We tested the ability of computer-assisted image analysis using convolutional neural networks (CNNs; a deep learning model for image analysis) to improve polyp detection, a surrogate of ADR.
METHODS:
We designed and trained deep CNNs to detect polyps using a diverse and representative set of 8,641 hand-labeled images from screening colonoscopies collected from more than 2000 patients. We tested the models on 20 colonoscopy videos with a total duration of 5 hours. Expert colonoscopists were asked to identify all polyps in 9 de-identified colonoscopy videos, which were selected from archived video studies, with or without benefit of the CNN overlay. Their findings were compared with those of the CNN using CNN-assisted expert review as the reference.
RESULTS:
When tested on manually labeled images, the CNN identified polyps with an area under the receiver operating characteristic curve of 0.991 and an accuracy of 96.4%. In the analysis of colonoscopy videos in which 28 polyps were removed, 4 expert reviewers identified 8 additional polyps without CNN assistance that had not been removed and identified an additional 17 polyps with CNN assistance (45 in total). All polyps removed and identified by expert review were detected by the CNN. The CNN had a false-positive rate of 7%.
CONCLUSION:
In a set of 8,641 colonoscopy images containing 4,088 unique polyps, the CNN identified polyps with a cross-validation accuracy of 96.4% and an area under the receiver operating characteristic curve of 0.991. The CNN system detected and localized polyps well within real-time constraints using an ordinary desktop machine with a contemporary graphics processing unit. This system could increase the ADR and decrease interval colorectal cancers but requires validation in large multicenter trials.
Copyright © 2018 AGA Institute. Published by Elsevier Inc. All rights reserved.
KEYWORDS:
Adenoma Detection Rate Improving Technology; Colorectal Cancer Prevention; Convolutional Neural Networks; Machine Learning
- PMID:
- 29928897
- PMCID:
- PMC6174102
- [Available on 2019-10-01]
- DOI:
- 10.1053/j.gastro.2018.06.037
놀라운 점은
이 논문이 컴퓨터공학 쪽에서 쓴 것 같은데
무려 Gastroenterology 지에 실렸다는 거...
미국 캘리포니아 어r바인에서 내서 그런 것일 수도 있지만...
8641 colonoscopy images (4,088 unique polyps) (이것은 수작업, 사람이 일일이 labeling) 로 training.
deep CNN architecture 로 VGG16, VGG19, ResNet50 을 사용함.
The CNN system detected and localized polyps well within real-time constraints using an ordinary desktop machine with a contemporary graphics processing unit.
All experiments were performed using modern TITAN X (Pascal) graphics processing units (NVIDIA, Santa Clara, CA) with 12 GB of RAM and a processing power of 11 TFLOPS.
Titan X 장착한 일반 데스크탑에서, real-time 으로 돌릴 수 있다고 주장... (타이탄 X 파스칼 (그래픽카드만 250만원이 넘음..) 장착한 데스크탑을 ordinary 라고 할 수 있는겨..??)
20개의 (총 5시간 짜리) 대장내시경 비디오를 보여주고, 폴립 찾아내라고 이 모델에게 시키고, human expert 에게도 찾아내라고 시키고.
CNN 은 96.4% 의 정확도로 찾아냄 (cross-validation accuracy of 96.4%)
CNN 의 false(+)는 7%.
어쨌든 CNN 사용해서, adenoma detection rate (ADR)를 확 더 끌어올림으로서,
일부(!.. expert) 사람 대장내시경 의사의 눈만큼 잘(?) adenoma를 놓치지 않고 잡아 낼 수 있다고...
어쨌든 놀라운 결과이고
내가 구상하던 아이디어가 논문화되어 나오는데 1년밖에 안걸리고 (그것도 의사들이 아닌 컴공에서 먼저 치고 나오는 구나...)
성능도 좋고...
의사들의 눈은 보고 판단하고 진단하는 관점에서, 앞으로 CNN architecture 에 확실이 도움을 받아야 하거나,
초심자 인간 의사들의 능력은 CNN architecture 들 보다 수준이 낮을 수도 있다는 것은, 위협적인 부분임.... 거꾸로 생각해서 이것을 이용하면 되고... AI (Augmented! Intelligence).
한편으로는.. diminutive small hyperplastic polyps 찾아내서 다 떼어내야 하는 것도 아닌데, 괜히 발견만 많이하고 시간도 많이 걸리고 비용도 많이 들게 되는 현실적인 문제들.
폴립 찾아낸 것 보면 대부분 Is, Isp 모양이고, 정작 놓치기 쉬운 IIa , IIb lesion 들이거나 LST type 은 아닌 것도 더 나아가야 할 부분..
그래도 찌꺼기도 있고, 주름도 다 안펴진 상태인데도 잘 구분해 내는 능력은 역시 CNN...
그리고 시행자에게는 detecting box로 힌트처럼 보여주기만 해도 찾는데 매우 도움이 되니까...
여기까지는 진단 (발견)에 도움을 주는 정도임.
폴립을 제거하는 것은 (치료)은 완전히 다른 문제.
치료로 들어가면, AI의 적용은 더 힘들고 어렵고 복잡하고, 로봇 공학이 발전하더라도 로봇이 폴립을 제거하는 날은 (특히 한국에서) 아직도 20-30년은 더 있어야 하지 않을까..
밥줄 걱정은 잠시 접어 둘 수 있는건가;;;
동영상 links : http://www.igb.uci.edu/colonoscopy/AI_for_GI.html
Supporting Material for the article "Deep Learning Localizes and Identifies Polyps in Real Time with 96% Accuracy in Screening Colonoscopy" (Urban et al., 2018, in press).
Examples of AI polyp predictions in the first set of 9 colonoscopy videos
Examples of AI polyp predictions in the second set of 11 colonoscopy videos
2018 Gastroenterology _ Deep Learning Localizes and Identifies Polyps in Real Time With 96% Acc in screening colonoscopy.pdf
