A session on artificial intelligence (AI) applications in biliary tract cancer (BTC) explored how AI-driven technologies are reshaping patient care in areas such as radiomics, prognostic modeling, and clinical decision-support systems.
Maria El Homsi, MD, began by highlighting the role of advanced AI-driven imaging techniques and radiomics in improving diagnosis, staging, and prognosis of cholangiocarcinoma (CCA). She explained that radiomics involves identifying regions of interest in medical images, extracting features, and combining these factors with clinical data using statistical methods to predict endpoints or outcomes. Current imaging modalities often fall short in diagnostic yield, whereas AI-driven radiomics enables a more accurate analysis of medical imaging data and integration with clinical information, advancing research and understanding of hepatobiliary diseases.
To illustrate the impact of radiomics, Dr El Homsi shared findings from a few recent studies. In 1 retrospective single-center study of 134 patients with intrahepatic CCA (iCCA) and hepatocellular carcinoma, radiomic models incorporating 15 imaging and 5 clinical features (chronic hepatitis, cirrhosis, and proteins alpha-fetoprotein, carcinoembryonic antigen, and carbohydrate antigen 19-9), achieved an area under the curve of 0.86 for radiomics alone and 0.89 for combined radiomics models, showing improved diagnostic accuracy.1 Another study assessing lymph node metastases in 100 patients with extrahepatic CCA using multiparametric magnetic resonance imaging reported high accuracy (90%), specificity (94%), and moderate sensitivity (75%).1 She also pointed to emerging applications of AI in detecting microvascular and neural invasion in iCCA, with additional prognostic capabilities demonstrated in studies for predicting early recurrence after curative resection and forecasting long-term survival outcomes. Key challenges in AI-based radiomics included issues with repeatability and reproducibility of features, time-consuming segmentation processes, and overfitting of models to specific datasets, limiting generalizability. Despite these limitations, Dr El Homsi concluded by urging clinicians to adopt AI, given its potential for improving imaging accuracy and advancing precision therapy.
Expanding the discussion, Changhoon Yoo, MD, PhD, focused on integrating AI into clinical practice and decision-making. He emphasized that BTC’s rarity, aggressive nature, and diagnostic complexity, combined with its rapidly evolving treatment options, create an urgent need to unify fragmented clinical, genomic, and proteomic data. Dr Yoo introduced AI tools such as Guardant360, Tempus AI, and CancerLinQ Suite, which support liquid biopsy analysis, pathology detection, and real-world data integration. Referencing the TOPAZ-1 trial, he demonstrated how adding durvalumab to gemcitabine/cisplatin therapy offers long-term survival benefits, with a median overall survival of 12.8 months compared with 11.5 months with gemcitabine/cisplatin alone (hazard ratio, 0.80; 95% confidence interval, 0.66-0.97; P=.021).2 He also highlighted AI-driven spatial analysis of tumor-infiltrating lymphocytes as a predictor of immune checkpoint inhibitor efficacy in BTC, citing a Korean multicenter where inflamed tumors responded better to anti–programmed death-ligand 1 therapy than excluded or desert phenotypes.
Dr Yoo stressed the need to utilize AI to enable greater levels of decentralized clinical trial enrollment for rare cancers like BTC, as geographic disparities often limit trial access. He concluded that AI could streamline real-world data integration, enhance prognosis, and deliver personalized treatment insights.
The session wrapped up with a dynamic discussion on the logistical and ethical challenges of integrating AI with electronic health records, feasibility of radiomics for early detection, and AI’s role in accelerating drug development. There was a general consensus that underscored the need for clinician oversight in AI-driven decision-making. Practical issues such as turnaround times for genomic and proteomic analyses were also raised, with Dr Yoo noting that proteomic results currently take up to 2 months. Overall, the session underscored AI’s transformative potential to improve diagnostic precision, enable personalized treatment decisions, and enhance clinical workflows in BTC care.
To sign up for our newsletter or print publications, please enter your contact information below.
