At the 2025 Annual Cholangiocarcinoma Foundation Conference, a working group that included co-chairs Vikram Deshpande and Benjamin Goeppert, as well as Julien Caldero, Guido Carpino, Matteo Fassan, Sanjay Kakar, Tim Kendall, David Klimstra, Beate Straub, Yoh Zen, Tiffany Mayer, and Fred Neubauer, discussed artificial intelligence (AI) usage in typing and identification of predictive markers for cholangiocarcinoma (CCA), ancillary testing in CCA, and biliary precursor lesions and etiology.
Vikram Deshpande first discussed machine learning for cancer pathology and the significance of the area under the curve (AUC). AUC assesses the performance of a machine learning tool; an AUC of 0.90 to 1.00 means it has excellent discrimination and is a highly accurate tool. He discussed 2 platforms in machine learning: the convolutional neural network (CNN) and the foundation model. CNN is an advanced AI model specifically designed for image analysis. This tool processes imaging in multiple layers: the input layer first accepts and processes the input image. The convolution layers then detect features using filters, and the activation layers highlight important information. Finally, the pooling layers reduce the complexity of the data. The fully connected layers combine features to make predictions, and the output layer provides the final prediction results.
One study that evaluated the utility of an AI algorithm on hepatobiliary cancer distinguished intrahepatic cholangiocarcinoma (iCCA) from colorectal metastasis. This study used 2 patient cohorts encompassing patients with iCCA and colorectal liver metastasis: an internal cohort (cohort 1; n=571 patients) recruited at Heidelberg University for training, validation, and internal testing and an external cohort (cohort 2; n=159 patients) curated at Mainz University Hospital serving as an independent test set to assess model generalizability. This AI tool used colors to distinguish iCCA and colon cancer from tissue biopsies. These were confirmed by AUC, which had AUCs of 0.956 from cohort 1 and 0.994 from cohort 2. Another study that evaluated an AI algorithm used deep learning-based phenotyping to reclassify combined hepatocellular cholangiocarcinoma (HCC). Patients had either HCC or iCCA, or both. The program was trained to detect and distinguish between HCC and CCA and performed well when it analyzed mixed tumors that contained both HCC and CCA, with AUC scores of 1 for one group of patients and 0.94 for another group of patients. The foundation model is a general-purpose AI model. One foundation model was trained on 4 million slides and can distinguish between many types of cancer. Machine learning can aid in diagnosis, predict mutations and fusions on hematoxylin and eosin (H&E) stained biopsies, predict prognosis on H&E, and predict response to therapy on H&E.
Vikram Deshpande then discussed ancillary testing in CCA. The first test he discussed was from a study that used novel branched DNA-enhanced albumin RNA in situ hybridization technology to diagnose iCCA. Albumin stains most HCCs of the small duct type but also stains other tumors including hepatoblastoma. One of the pitfalls for pathologists is the tendency to order many stains and deplete the biopsy sample. Some guidelines for pathologists to follow are to utilize targeted, limited immunohistochemistry stains when necessary and to prioritize molecular testing and tissue preservation.
Lastly, Sanjay Kakar discussed precursor lesions and etiology. It is important to study biliary precursor lesions because it is important to detect these lesions early and understand their molecular pathogenesis. This can help to develop targeted therapies and clarify tumor evolution and phylogeny. He emphasized that CCA at the molecular level is heterogeneous, and it is important to study the precursor lesions of CCA that arise in different settings because the molecular alterations are known to be different.
In summary, this conference brought together leading experts to discuss cutting-edge advancements in AI, ancillary testing, and the study of biliary precursor lesions in CCA. The discussions highlighted the transformative potential of AI in cancer pathology, particularly in distinguishing tumor types, predicting mutations, and aiding in therapy selection with high accuracy as demonstrated by advanced platforms like CNN and foundation models. Ancillary testing advancements, such as RNA in situ hybridization, underscore the importance of preserving biopsy samples while achieving accurate diagnoses. Additionally, the study of biliary precursor lesions and their molecular heterogeneity offers a pathway to early detection, targeted therapies, and a deeper understanding of tumor evolution. These insights emphasize the need for continued collaboration, innovation, and translational research to improve CCA diagnosis, treatment, and patient outcomes.
Goeppert B, Deshpande V. WG8-Pathology. Presented at: 2025 Annual Cholangiocarcinoma Foundation Conference. April 9-11, 2025; Salt Lake City, UT.
To sign up for our newsletter or print publications, please enter your contact information below.
