Convolutional neural network detects malignant cholangioscopy image features
Click Here to Manage Email Alerts
LAS VEGAS – A trained convolutional neural network detected key cholangioscopy image features suggestive of malignancy supporting intra-procedural decision making, according to a presentation at the ACG Annual Scientific Meeting.
“Indeterminate biliary strictures remain a diagnostic challenge despite advancements in radiologic, endoscopic and laboratory testing,” Bachir Ghandour, MD, a postdoctoral research fellow at Johns Hopkins University Hospital, said. “It has been reported that up to 25% of patients presumed to have malignant strictures during cholangioscopy show benign pathology after undergoing major surgical intervention. Interpretation of the visual findings during cholangioscopy remains challenging even for experienced endoscopists.”
Using 1,371,605 cholangioscopy images and clinical data from 528 patients who underwent cholangioscopy for the evaluation of indeterminate biliary strictures, Ghandour and colleagued aimed to develop a software tool that classifies strictures as either benign or malignant. They annotated images for abnormal features of papillary masses, dilated and tortuous vessels and ulceration suggestive of malignancy and trained a convolutional neural network (CNN) based on ResNet-18 to detect abnormality presence. External validation included a training set of 254 patients and a test set of 95 patients.
Among 217 malignant biliary strictures and 132 benign biliary strictures, researchers identified papillary masses in 66.3% and 23.5% of patients, dilated and tortuous vessels in 36.4% and 7.6% of patients and ulceration in 23.5% and 7.6% of patients.
In detecting abnormal images, the CNN had a sensitivity of 0.81 (95% CI, 0.72-0.91), specificity of 0.91 (95% CI, 0.86-0.97), positive predictive value of 0.93 (95% CI, 0.88-0.98), negative predictive value of 0.77 (95% CI, 0.66-0.88) and area under the curve of 0.86 (95% CI, 0.8-0.92).
“What we’ve managed so far is to develop and validate a machine learning software tool that can identify abnormal cholangioscopy images with features suggestive of malignancy and differentiate them from normal biliary mucosa,” Ghandour concluded. “This shows promise for our final software that can combine cholangioscopy visual findings with patient’s clinical data to predict malignant nature of indeterminate biliary strictures.”
Collapse
Ghandour B, et al. Abstract: Machine learning for classification of indeterminate biliary strictures during cholangioscopy. Presented at: ACG Annual Scientific Meeting; Oct. 22-27, 2021; Las Vegas (hybrid meeting).
Read more about
Click Here to Manage Email Alerts