Computer-aided diagnosis of serrated colorectal lesions using non-magnified white-light endoscopic images

Abstract

Purpose: Computer-aided diagnosis systems for polyp characterization are commercially available but cannot recognize subtypes of sessile lesions. This study aimed to develop a computer-aided diagnosis system to characterize polyps using non-magni ed white-light endoscopic images.Methods: A total of 2249 non-magni ed white-light images from 1030 lesions including 534 tubular adenomas, 225 sessile serrated adenoma/polyps and 271 hyperplastic polyps in the proximal colon were consecutively extracted from an image library and divided into training and testing datasets (4:1), based on the date of colonoscopy. Using ResNet-50 networks, we developed a classi er (1) to differentiate adenomas from serrated lesions, and another classi er (2) to differentiate sessile serrated adenoma/polyps from hyperplastic polyps. Diagnostic performance was assessed using the testing dataset. The computer-aided diagnosis system generated a probability score for each image, and a probability score for each lesion was calculated as the weighted mean with a log 10 -transformation. Two experts (E1, E2) read the identical testing dataset with a probability score.Results: The area under the curve of classi er (1) for adenomas was equivalent to E1 and superior to E2 (classi er 86%, E1 86%, E2 69%; classi er vs. E2, p<0.001). In contrast, the area under the curve of classi er (2) for sessile serrated adenoma/polyps was inferior to both experts (classi er 55%, E1 68%, E2 79%; classi er vs. E2, p<0.001). Conclusion:The classi er (1) developed using white-light images alone compares favorably with experts in differentiating adenomas from serrated lesions. However, the classi er (2) to identify sessile serrated adenoma/polyps is inferior to experts.

A colonial serrated polyp classification model using white-light ordinary endoscopy images with an artificial intelligence model and TensorFlow chart

“…It is worth noting that our method allows the endoscopist to build/use the model with a limited sample size using white light images from ordinary endoscopy, not NBI or magnified images. Our AI model with only white-light images outperform the SSA and TA classification model build in Nemoto et al [ 10 ], which obtain accuracy ranges from 77 to 87%. This might be due to the fact we combine the data argumentation with deep-learning model.…”

“…Therefore, using enhanced images to build an AI classification model can significantly limit the generality of the model. Nemoto et al [ 10 ] built an AI model that can distinguish SSA and TA using white-light images. Their accuracy ranges from 77 to 87%.…”

“…In this study, the polyp classification model should be embedded in the endoscopy system, and an instant classification result is needed during the endoscopy section. Furthermore, Nemoto et al [ 10 ] use ResNet50 as their modeling backbone in distinguish the SSA and HP and they find that their AI model with ResNet50 cannot outperform the trained physicians. Therefore, the ResNet model might not be the best fit for classifying the colonic polyps.…”

Impact of computer‐aided characterization for diagnosis of colorectal lesions, including sessile serrated lesions: Multireader, multicase study

“…A previous study showed the difficulty of developing a CADx system able to distinguish SSLs from hyperplastic polyps. Nemoto et al 33 reported that their developed CADx showed a smaller area under the receiver operating characteristic curve (CADx vs. expert: 55% vs. 68-79%). This complication may be caused by interobserver variability among pathologists.…”

Color-coded parametric imaging support display of vessel hemorrhage—an in vitro experiment and clinical validation study

“…In this study, we initially selected 8 parameters that may affect the performance of DSA images (14,15): (1) the speed of contrast agent injection, (2) the timing of contrast agent injection, (3) the dose of x-ray, (4) the DSA imaging frame rate, (5) the pressure of contrast agent injection, (6) the flow velocity, 6) flow rate, (7) number of broken fibers (the number of cutting fibers of the dialyzer to simulate bleeding) and ( 8) the distance of contrast agent injection from the injection point to the dialyzer. In order to obtain a sufficient number of experiments, we used an orthogonal experimental design to create 27 experimental groups (Supplementary Material S2).…”