Powers, Patrick

ABSTRACT Background Sampling techniques have poor accuracy for classifying biliary strictures as benign or malignant. Previously, a cholangioscopy artificial intelligence (AI) outperfromed sampling techniques based solely on analysis of previously recorded cholangioscopy footage. The aim of this trial was to compare the performance of a real-time cholangioscopy AI to both sampling techniques and human observers for the task of biliary stricture classification. Methods A cholangioscopy AI computer connected directly to a cholangioscope console. The computer analyzed the cholangioscopy video stream during procedures for suspected biliary strictures. The primary outcome of the study was comparison of the performance of cholangioscopy AI to sampling techniques – brush cytology and transpapillary forceps biopsy – for biliary stricture classification. Secondary outcomes included comparison of the AI classification performance to that of human observers (separated into junior-level and experienced-level cohorts) who reviewed the cholangioscopy footage. Results A total of 41 patients were enrolled in the trial and had biliary strictures analyzed by cholangioscopy AI. For the classification of strictures, the AI had greater classification accuracy than standard sampling techniques (87.8% versus 67.4%; p = 0.043). Additionally, the cholangioscopy AI was significantly more accurate for biliary stricture classification than both junior-level (87.8% versus 61.5%; p = 0.001) and experienced endoscopists (87.8% versus 63.15%; p = 0.011). Conclusions This trial demonstrates that sampling techniques and human assessment of biliary strictures are flawed and there may be a benefit to the use of a cholangioscopy AI system to aid in biliary stricture classification.

Abstract Clinicians struggle to accurately classify biliary strictures as benign or malignant. Current endoscopic retrograde cholangiopancreatography (ERCP)-based sampling modalities including brush cytology and forceps biopsy have poor sensitivity for pathologic confirmation of malignancy. Cholangioscopy allows for direct visualization and sampling of biliary pathology; however, this technology is also associated with inaccurate classification of biliary disease. Previously, an artificial intelligence (AI) system that analyzes cholangioscopy footage was found to be more accurate in diagnosing biliary malignancy than ERCP sampling techniques. The aim of this study was to validate this AI system on a new series of examinations. Three academic centers collected all available unedited cholangioscopy recordings. The videos were processed by the cholangioscopy AI system. After analyzing videos, the AI system provided predictions as to whether malignancy was present. AI performance in classifying strictures was compared with the performance of brush cytology and forceps biopsy. 112 cholangioscopy examinations (containing 4 817 081 images) were generated from 99 patients. Of those examinations, 61 (54.5%) were for investigation of biliary strictures (31 [50.8%] benign, 30 [49.2%] malignant). For the correct classification of strictures, the AI system was 80.0% sensitive and 90.3% specific. It was also significantly more accurate for stricture classification (85.2%) than brush cytology (52.5%; P<0.001), forceps biopsy (68.2%; P=0.04), and the combination of brush cytology and forceps biopsy (66.7%; P=0.02). A previously developed cholangioscopy AI system was found to continually outperform standard ERCP sampling modalities for accurate identification of malignancy, without additional retraining, in a multicenter validation cohort.

Abstract Background The authors previously developed an artificial intelligence (AI) to assist cytologists in the evaluation of digital whole‐slide images (WSIs) generated from bile duct brushing specimens. The aim of this trial was to assess the efficiency and accuracy of cytologists using a novel application with this AI tool. Methods Consecutive bile duct brushing WSIs from indeterminate strictures were obtained. A multidisciplinary panel reviewed all relevant information and provided a central interpretation for each WSI as being “positive,” “negative,” or “indeterminate.” The WSIs were then uploaded to the AI application. The AI scored each WSI as positive or negative for malignancy (i.e., computer‐aided diagnosis [CADx]). For each WSI, the AI prioritized cytologic tiles by the likelihood that malignant material was present in the tile. Via the AI, blinded cytologists reviewed all WSIs and provided interpretations (i.e., computer‐aided detection [CADe]). The diagnostic accuracies of the WSI evaluation via CADx, CADe, and the original clinical cytologic interpretation (official cytologic interpretation [OCI]) were compared. Results Of the 84 WSIs, 15 were positive, 42 were negative, and 27 were indeterminate after central review. The WSIs generated on average 141,950 tiles each. Cytologists using the AI evaluated 10.5 tiles per WSI before making an interpretation. Additionally, cytologists required an average of 84.1 s of total WSI evaluation. WSI interpretation accuracies for CADx (0.754; 95% CI, 0.622–0.859), CADe (0.807; 95% CI, 0.750–0.856), and OCI (0.807; 95% CI, 0.671–0.900) were similar. Conclusions This trial demonstrates that an AI application allows cytologists to perform a triaged review of WSIs while maintaining accuracy.

Objective The diagnosis of autoimmune pancreatitis (AIP) is challenging. Sonographic and cross-sectional imaging findings of AIP closely mimic pancreatic ductal adenocarcinoma (PDAC) and techniques for tissue sampling of AIP are suboptimal. These limitations often result in delayed or failed diagnosis, which negatively impact patient management and outcomes. This study aimed to create an endoscopic ultrasound (EUS)-based convolutional neural network (CNN) model trained to differentiate AIP from PDAC, chronic pancreatitis (CP) and normal pancreas (NP), with sufficient performance to analyse EUS video in real time. Design A database of still image and video data obtained from EUS examinations of cases of AIP, PDAC, CP and NP was used to develop a CNN. Occlusion heatmap analysis was used to identify sonographic features the CNN valued when differentiating AIP from PDAC. Results From 583 patients (146 AIP, 292 PDAC, 72 CP and 73 NP), a total of 1 174 461 unique EUS images were extracted. For video data, the CNN processed 955 EUS frames per second and was: 99% sensitive, 98% specific for distinguishing AIP from NP; 94% sensitive, 71% specific for distinguishing AIP from CP; 90% sensitive, 93% specific for distinguishing AIP from PDAC; and 90% sensitive, 85% specific for distinguishing AIP from all studied conditions (ie, PDAC, CP and NP). Conclusion The developed EUS-CNN model accurately differentiated AIP from PDAC and benign pancreatic conditions, thereby offering the capability of earlier and more accurate diagnosis. Use of this model offers the potential for more timely and appropriate patient care and improved outcome.