World Congress at ACG2017

Presidential Plenary Session 1 (Free Paper/Abstract Presentations)

4 - A Prospective Validation of Deep Learning for Polyp Auto-detection during Colonoscopy

Monday, October 16
8:36 AM - 8:48 AM
Location: Valencia Ballroom (Level 4)



Award: 2017 International Award

Category: Colorectal Cancer Prevention       

Pu Wang, MD1, Xiao Xiao, PhD2, Jingjia Liu, BSc2, Liangping Li, MD1, Mengtian Tu, MS1, Jiong He, BSc2, Xiao Hu, MS1, Fei Xiong, MS1, Yi Xin, BSc2, Xiaogang Liu, MD1
1Sichuan Academy of Medical Sciences & Sichuan Provincial People’s Hospital, Chengdu, Sichuan, China (People's Republic); 2Shanghai Wision AI Co., Ltd, Shanghai, China (People's Republic)
Introduction: Colonoscopy screening can reduce the risk of death from colorectal cancer through detection of precancerous adenomas. However, there is a significant adenoma miss rate ranging from 6-27% depending on a variety of polyp and operator characteristics. Unrecognized polyps within the visual field have been very difficult to address. Deep learning is a computational method that is very successful in automatic object detection in images and videos. But the feasibility of applying deep learning for medical imaging detection in the clinical setting has never been determined. Thus, automatic polyp detection has received tremendous attention.

Methods: We built a deep learning algorithm for automatic polyp detection using a retrospective set of 5,545 endoscopists-annotated images from colonoscopies performed between 2007 and December 2015. The resultant algorithm was validated on a large-scale prospective clinical set of 27,461 images from 1,291 colonoscopies performed between January and December 2016. A panel of 5 national licensed endoscopists, confirmed by colonoscopy and pathological reports, assessed the algorithm’s detection results.

Results: The prospective validation set consisted of 27,461 colonoscopy images from 1,235 patients (mean age, 57.4; 36.7% women; 1,495 pathological confirmed polyps). In the free-response ROC analysis, the algorithm had an AUC of 0.958 (95% CI, 0.954-0.962). Using the high sensitivity operating point, the algorithm had a sensitivity of 94.96% at a false positive rate of 7.76% (specificity of 92.01%). Using the low false positive rate (i.e. high specificity) operating point, the algorithm had a sensitivity of 92.35% at a false positive rate of 2.98% (specificity of 97.05%). For subgroups of flat polyps (Yamada type I), smaller polyps (<=0.5cm), and isochromatic polyps, the algorithm had an AUC of 0.943 (95% CI, 0.936-0.950), 0.957 (95% CI, 0.952-0.962), and 0.957 (95% CI, 0.952-0.962), respectively.

Discussion: This large-scale prospective validation of deep learning applied for automatic polyp detection during colonoscopy showed high sensitivity and low false positive rate (i.e. high specificity) in the clinical setting. The reporting time of this algorithm was within 60-80ms, which is ready for real-time assist with polyp detection during colonoscopy.

Supported by Industry Grant: No














































































































































































































Characteristics


Developmental Data Set Validation Data Set
Patient Demographics    
 No. of unique individuals

1,290 (100)



1,235 (100)


 Age, mean (SD), y

NA



57.35 (12.85)


 Female, No. (%)

NA



451 (36.68)



Image Characteristics Distribution


   

  Total No. of images (%)



5,545 (100)



27,461 (100)



  No. of images with at least one polyp (%)



3,634 (65.5)



5,827 (21.22)



  No. of images with only pathological confirmed polyp (%)



NA



5,556 (20.2)



  No. of images with only flat polyp (%)



NA



2,106 (7.67)



  No. of images with only smaller polyp (%)



NA



3,830 (13.95)



  No. of images with only isochromatic polyp (%)



NA



3,147 (11.46)



  No. of images with obscured polyp (%)



NA



1,918 (6.98)



  No. of images with polyp on the edge of visual field(%)



NA



217 (0.79)



  No. of images with blurred polyp (%)



NA



2,837 (10.33)



  No. of images with polyp in distance (%)



NA



906 (3.3)



  No. of images with polyp in dark (%)



NA



1,268 (4.62)



Polyp Characteristics Distribution



 


 

  Total No. of colorectal polyps with biopsy* (%)



NA



1,495 (100)



Location Distribution



 


 

  Rectum (%)



NA



625 (41.81)



  Sigmoid colon (%)



NA



195 (13.04)



  Descending colon, including splenic flexure (%)



NA



108 (7.22)



  Transverse colon (%)



NA



253 (16.92)



  Ascending colon, including hepatic flexure (%)



NA



251 (16.79)



  Ileocecum (%)



NA



63 (4.21)



Pathological Feature Distribution



 


 

  Carcinomatous (%)



NA



57 (3.81)



  Adenomatous (%)



NA



1,044 (69.83) 



  Hyperplastic (%)



NA



200 (13.38)



  Inflammatory (%)



NA



184 (12.31)



  Hamartomatous (%)



NA



6 (0.4)



  Others** (%)



NA



4 (0.27)



* A few typical polyps were not taken biopsy because of patients refusal, or risk factors including bad coagulation function or low platelet, et al.


   

** Spindle cell tumor x 4


   
     
     
     
     
     

Deep Learning Architecture: Our detection algorithm is a deep convolutional neural network (CNN): Cambridge's SegNet CNN architecture with the last 4 layers deleted from the encoder network, and made the corresponding modification to the decoder network. Data flow is from left to right: a colonoscopy image is sequentially warped into an overlaid image highlighting the detected polyp. This CNN is randomly initialized and trained on the carefully selected training data of 4,495 endoscopic images, of which 2,607 had at least one colorectal polyp satisfying the sub-feature criteria with prominent protuberance polyps excluded. (http://mi.eng.cam.ac.uk/projects/segnet/)
Validation Set Performance for Automatic Polyp Detection: A. Algorithm performance on all-polyp group (red curve) and polyp-with-biopsy group (blue dotted curve). The algorithm had an AUC of 0.958 (95% CI, 0.954-0.962). At the low false positive rate (i.e. high specificity) operating point, the algorithm had a sensitivity of 92.35% at a FPR of 2.98% (specificity of 97.05%). At the high sensitivity operating point, the algorithm had a sensitivity of 94.96% at a FPR of 7.76% (specificity of 92.01%).
B. Algorithm performance on flat-polyp subgroup (blue dotted curve), compared with all-polyp group (red curve). The algorithm achieved an AUC of 0.943 (95% CI, 0.936-0.950). At the low false positive rate (i.e. high specificity) operating point, the algorithm had a sensitivity of 88.39% at a FPR of 3.08% (specificity of 97.05%). At the high sensitivity operating point, the algorithm had a sensitivity of 92.41% at a FPR of 8.24% (specificity of 92.01%).
C. Algorithm performance on smaller-polyp subgroup (blue dotted curve), compared with all-polyp group (red curve). The algorithm achieved an AUC of 0.957 (95% CI, 0.952-0.962). At the low false positive rate (i.e. high specificity) operating point, the algorithm had a sensitivity of 92.17% at a FPR of 2.99% (specificity of 97.05%). At the high sensitivity operating point, the algorithm had a sensitivity of 94.81% at a FPR of 7.96% (specificity of 92.01%).
D. Algorithm performance on isochromatic-polyp subgroup (blue dotted curve), compared with all-polyp group (red curve). The algorithm achieved an AUC of 0.957 (95% CI, 0.952-0.962). At the low false positive rate (i.e. high specificity) operating point, the algorithm had a sensitivity of 91.96% at a FPR of 2.99% (specificity of 97.05%). At the high sensitivity operating point, the algorithm had a sensitivity of 94.88% at a FPR of 8.03% (specificity of 92.01%).
Demonstration of Algorithm’s Detection Results on All-polyp Group: A. For incomplete polyp presentation caused by occlusions and visual field, the algorithm’s detection results were sensitive and accurate.
B. For polyps in dark and far distance, the algorithm’s detection results were sensitive and accurate.
C. For polyps close to water bubbles, wound, collapsed mucosa, light spots, etc. the algorithm’s detection results were sensitive and accurate.
D. For colonoscopy images with obvious motion blur, the algorithm’s detection results were sensitive and accurate.

Citation: . A PROSPECTIVE VALIDATION OF DEEP LEARNING FOR POLYP AUTO-DETECTION DURING COLONOSCOPY. Program No. 4. World Congress of Gastroenterology at ACG2017 Meeting Abstracts. Orlando, FL: American College of Gastroenterology.

Pu Wang

Sichuan Academy of Medical Sciences & Sichuan Provincial People’s Hospital
Chengdu, Sichuan, China (People's Republic)

Presentation(s):

Send Email for Pu Wang


Assets

4 - A Prospective Validation of Deep Learning for Polyp Auto-detection during Colonoscopy



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for A Prospective Validation of Deep Learning for Polyp Auto-detection during Colonoscopy