Optical diagnosis is the in-vivo prediction of colorectal polyp histopathology. Inter-observer variability amongst endoscopists has limited its application in clinical practice. Artificial intelligence may lead to a new generation of clinical support tools capable of characterising polyps. Research in this field has often relied upon retrospective datasets, which are subject to ... [Show full abstract] sample selection bias, and consist of a limited number of images of each polyp.Our aim was to develop a convolutional neural network (CNN) to characterise colorectal polyps as adenomatous or non-adenomatous using data collected prospectively.
Video data was collected prospectively from colonoscopy procedures at a single centre using Olympus 260 and 290 series scopes. Histopathological classification, location and morphology was recorded for each polyp.Video sequences of polyps in Narrow Band Imaging (NBI) and NBI-Near Focus (NBI-NF) were extracted. Both imaging modalities were used to increase the generalisability of the CNN. Frames with poor visualisation of the polyp surface texture due to mucus, stool, halation or motion artifact were excluded. The ground truth for each frame was the polyp annotated with a bounding box and labelled with the histopathology.
A ResNet-101 CNN pre-trained on ImageNet was developed to classify the visual appearance of colorectal polyps as adenomatous or non-adenomatous.
The final dataset consisted of 371 histologically confirmed polyps (235 adenomas, 77 sessile serrated lesions, 58 hyperplastic, 1 traditional serrated adenoma) from 199 patients with a total of 31,110 video frames annotated. Data was split, as shown in Figure 1, into a training (~50%), validation (~10%), and testing dataset (~40%) with no overlap of polyps or patients.On a per-frame analysis, the accuracy of the CNN optical characterisation was 91%, with a sensitivity of 91% to diagnose adenomas and a specificity of 90%. The CNN achieved an area under the curve (AUC) of 97%. On a per polyp analysis, the accuracy of the CNN characterisation was 92%, with a sensitivity of 92% and a specificity of 93%.
View this table:
• View inline
• View popup
• Download powerpoint
Abstract HTU-1 Table 1
The largest annotated dataset of NBI polyp images has been collated for the training and evaluation of artificial intelligence to support optical diagnosis. This work demonstrated the capability of AI to differentiate adenomatous from non-adenomatous polyps in-vitro, with a high level of accuracy.