Category: Sunday Poster Session
P93 - Gastroenterology Fellow Level Performance on Colonoscopy Frame Classification With Deep Neural Networks
Sunday, Oct 15
3:30 PM – 7:00 PM
Lei Zhang, PhD1, Zongwei Zhou, BSc1, Hassan Siddiki, MD, MS2, Naveen Sai Madiraju, BS1, Noemi Baffy, MD, MPH3, Diana Franco, MD3, Francisco C. Ramirez, MD3, Suryakanth R. Gurudu, MD3, Jianming Liang, PhD1
1Arizona State University, Scottsdale, AZ; 2Mayo Clinic College of Medicine, Scottsdale, AZ; 3Mayo Clinic, Scottsdale, AZ
Introduction: Colorectal cancer (CRC) is second most cause of cancer related deaths in USA. Colonoscopy is the preferred technique for CRC screening and prevention. However, a colonoscopy is an operator dependent procedure, wherein human factors, e.g., skill of the endoscopist and his/her concentration during the intervention, can lead to missed detection of polyps. To address this problem, we propose a deep-learning based automatic quality assessment method that provides feedback to an endoscopist during the colonoscopy. The proposed method has been compared against three gastroenterology fellows and can achieve an acceptable level of accuracy.
Methods: 760 frames are selected from 10 videos and 21 patches of size 256 * 256 are extracted from each frame with label of good quality, poor quality, or ambiguous (Fig. 1). Two gastroenterology attending readers in a blinded manner analyzed each frame independently and each ambiguous frame was analyzed twice. Both readers have > 10 years experience and Adenoma Detection Rate (ADR) is > 40%. Ground truth is determined when 2 readers agree. The Residual network (ResNet) is used as backbone of convolutional neural network (CNN). The CNN was fine-tuned from pre-trained ResNet on ImageNet dataset. We performed 10-fold cross-validation to test against three gastroenterology fellows’ performance.
Results: Receiver Operating Characteristic (ROC) curve is used to evaluate performance of the CNN and the gastroenterology fellow (Fig. 2). The CNN achieves superior performance to a gastroenterology fellow if a (1-specificity, sensitivity) point of the gastroenterology fellow lies below the ROC curve. The blue points represent the average of the gastroenterology fellows, with error bars denoting the standard deviation. In the ROC, the area under the curve (AUC) is a depiction of the CNN performance. The larger the AUC represents the better performance of the CNN. The CNN outperforms the average performance of the gastroenterology fellows on all three classes colonoscopy frame classification and is on par with all gastroenterology fellows.
Discussion: The CNN is on par with all gastroenterology fellows on the colonoscopy frame classification, demonstrating an artificial intelligence capable of classifying colonoscopy frames with a level of competence comparable to a gastroenterology fellow. Our study generates feasibility of an image analysis framework that can be employed to automatically interpret the colonoscopy video quality.
Supported by Industry Grant: No
Figure 1: Representative frames of good quality, poor quality, and ambiguous.
Figure 2: Colonoscopy frame classification performance of the CNN and the gastroenterology fellows. A gastroenterology fellow provides a predication per frame represented by a green point. The blue points represent the average of the gastroenterology fellows, with error bars denoting the standard deviation. The CNN achieves superior performance to a gastroenterology fellow if a (1-specificity, sensitivity) point of the gastroenterology fellow lies below the ROC curve.
Citation: . GASTROENTEROLOGY FELLOW LEVEL PERFORMANCE ON COLONOSCOPY FRAME CLASSIFICATION WITH DEEP NEURAL NETWORKS. Program No. P93. World Congress of Gastroenterology at ACG2017 Meeting Abstracts. Orlando, FL: American College of Gastroenterology.