April 01, 2025
The gold standard diagnostic modality for diagnosing malignant cholangiocarcinoma is cytologic evaluation of bile duct brushing specimens. However, despite having a high diagnostic specificity, bile duct brushing cytologic analysis has some limitations. Manual interpretation is time-consuming and has a low diagnostic sensitivity, a low negative predictive value and high interobserver variability. Although the diagnostic sensitivity can be improved by combining bile duct brushing cytology with tumor marker analysis and other testing, including fluorescence in situ hybridization (FISH) and molecular testing, researchers also are studying the use of artificial intelligence (AI) tools.
Previous studies have established that the use of convolutional neural networks can improve diagnostic sensitivity when performing specific diverse image analysis tasks such as classifying lesions identified on cholangioscopy and classifying colon polyps. In a publication in Clinical Gastroenterology and Hepatology in 2024, Mayo Clinic researchers and others described their initial experience in training and validating an AI tool designed to assist in the evaluation of digital whole-slide images (WSIs) generated from bile duct brushing specimens.
The AI tool that the researchers created was equipped with computer-aided detection (CADe) and computer-aided diagnosis (CADx) components. The CADe component was designed to identify and prioritize the subregions, also called tiles, within a WSI where malignancy was most likely to be present. The CADx component was designed to autonomously differentiate malignant from benign bile duct brushing WSIs. The results from the initial study of the AI tool demonstrated that expert cytologist review, CADe review and CADx review performed with similar accuracy when classifying bile duct brushing WSIs.
After developing the AI tool, the researchers also created a clinical web-based application that allows cytologists to use the AI tool when making diagnostic assessments of bile duct brushing WSIs. The web‐based application uses the CADe component to prioritize WSI review on the basis of the AI‐determined malignancy risk. The system is designed to present cytologists with the tiles most concerning for malignancy for review before they provide an interpretation for the entire WSI.
In a pilot study published in Cancer Cytopathology in 2024, the team examined the accuracy and efficiency of cytologists using the AI-enhanced web-based application to analyze a new prospective biliary WSI dataset.
Methods
The study included a total of 84 slides. The researchers compared the diagnostic performance of the cytologic interpretation conducted by:
- The original, clinical cytologic interpretation rendered by cytologists with access to pertinent clinical information (such as patient history, tumor marker levels, results of FISH and prior cytologic sampling) but without AI assistance.
- Four blinded expert cytologists using the AI-enhanced web-based application (CADe).
- The AI, without cytologists, which autonomously prioritized cytologic tiles by the likelihood that malignant material was present and scored each WSI as positive or negative for malignancy (CADx).
A multidisciplinary panel with access to all clinical, laboratory, pathology and imaging data related to the WSIs served as the gold standard diagnosis to determine the performance characteristics of the review methods listed above.
Results
Overall, the researchers note that the pilot study demonstrates that an AI application equipped with both CADx and CADe elements allows cytologists to perform a triaged review of WSIs efficiently while maintaining accuracy. Cytologists performing the AI-assisted review spent significantly less time evaluating WSIs that the AI had scored as being less concerning for malignancy.
Pilot study data
"Artificial intelligence is poised to transform healthcare, offering enhanced capabilities, reliability and efficiency similar to other industries," explains Rondell P. Graham, M.B.B.S. Dr. Graham, the corresponding author on the pilot study publication, is a pathologist and researcher at Mayo Clinic in Rochester, Minnesota. "Our team demonstrated that a neural network can automatically interpret bile duct brushings at a much faster speed and triage cases for interpretation with performance characteristics that are noninferior to experts. This sets the stage for the development of automated workflows and provides a framework for AI to enhance both diagnostics and operational efficiency."
Dr. Graham and co-authors say that the study findings suggest that AI can play a role in the future of lab medicine practices, including cytopathology. "Our profession will need to improve capacity and address the labor shortage," explains Dr. Graham. "Versions of this tool could address these challenges by supplementing the team in the prescreening role. And the almost eightfold faster interpretation of cases enabled by the model could address the capacity limitations affecting cytopathology practices and groups."
Dr. Graham notes that these findings need to be validated on larger patient cohorts from multiple institutions and using varied slide preparations. "Subsequent implementation pilots will provide additional insight into how to implement this and other AI-enabled tools in practice to benefit patients," says Dr. Graham.
For more information
Marya NB, et al. Development of a computer-aided prediction tool for evaluating brushing samples of biliary strictures. Clinical Gastroenterology and Hepatology. 2024;22:185.
Marya NB, et al. Utilization of an artificial intelligence-enhanced, web-based application to review bile duct brushing cytologic specimens: A pilot study. Cancer Cytopathology. 2024;132:779.
Refer a patient to Mayo Clinic.