Computer-Aided Characterisation Of Colorectal Polyps Using Artificial Intelligence

Aims Optical diagnosis is the in-vivo prediction of colorectal polyp histopathology but inter-observer variability amongst endoscopists has limited its application in clinical practice. Artificial Intelligence, using deep learning, may lead to a new generation of clinical support tools capable of characterising polyps. Our aim was to develop a convolutional neural network (CNN) to characterise colorectal polyps as adenomatous or non-adenomatous.

Methods Data was collected from unaltered colonoscopy videos from 8 endoscopists at a single centre using Olympus 260 and 290 series scopes. Histopathological classification was recorded for each polyp.

The dataset was created using Narrow Band Imaging (NBI) and NBI-Near Focus (NBI-NF) video sequences. Frames with limited visualisation of the polyp surface texture were excluded. The remaining frames were annotated with bounding boxes around polyps and labelled with the histopathology. The annotations were referenced as the gold standard.

A ResNet-101 CNN pre-trained on ImageNet was developed to classify the visual appearance of colorectal polyps as adenomatous or non-adenomatous. During inference, the probability scores computed by the CNN were used as confidence for its prediction. A score above 70 % was defined as a confident polyp characterisation of adenomatous, below 30 % as non-adenomatous, and 30-70 % as a low confidence characterisation.

Results The dataset consisted of 187 polyps (122 adenomas, 48 sessile serrated lesions, 17 hyperplastic) from 71 patients with a total of 41,171 frames. Data was split into a training (~65 %), validation (~5 %), and testing set (~30 %) with no overlap of data/patients.

The CNN achieved a confident diagnosis in 84 % of frames in the test set. On a per-frame analysis, excluding low confident diagnoses, the CNN diagnosed adenomas with a sensitivity of 92 % and specificity of 90 %. On a per-polyp analysis, sensitivity was 93 % and specificity 87 %. The area under the curve was 96 %.

Tab. 1

	Training Dataset	Validation Dataset	Testing Dataset
Number of patients	38	7	26
Number of adenomas polyp (frames)	75 (22,007)	6 (1,143)	41 (8,218)
Number of non-adenomatous polyps (frames)	40 (5,412)	2 (788)	23 (3,603)
Total number of frames	27,419	1,931	11,821

Conclusions The CNN achieved promising results to differentiate adenomatous from non-adenomatous polyps.

Citation Kader R, Brandao P, Ahmad O et al. eP479 COMPUTER-AIDED CHARACTERISATION OF COLORECTAL POLYPS USING ARTIFICIAL INTELLIGENCE. Endoscopy 2021; 53: S255.

Publication History

Article published online:
19 March 2021