News Research Gastrointestinal Cancers Decision-Making Support

ASTRA Agentic System Improves Accuracy of Resectability Assessment in Perihilar Cholangiocarcinoma

November 13, 2026 Lisa Astor 5 min read
Share Share via Email Share on Facebook Share on LinkedIn Share on Twitter

An agentic system was developed to support preoperative clinical decision-making for patients with perihilar cholangiocarcinoma. The multistep evaluation and validation process was presented at the ESMO Congress 2026 (Abstract 172R0).  

“ASTRA demonstrates that an agentic AI system can potentially augment multidisciplinary decision-making in perihilar cholangiocarcinoma, improving both assessment accuracy and workflow efficiency,” said presenting author Yanqi Huang, MD, an attending radiologist at Guangdong Provincial People’s Hospital, Southern Medical University, and the Guangdong Provincial Key Laboratory of AI in Medical Image Analysis and Application, both in Guangzhou, China. 

Decision-making in perihilar cholangiocarcinoma requires a multidisciplinary approach plus the synthesis of radiologic findings, clinical data, and guideline-based resectability criteria. Dr. Huang explained that there is an unmet need for a more standardized process to generate multidisciplinary team briefs. Because this process involves multiple steps and guideline-based tasks, an agentic system was considered well-suited to it.  

System and Study Design

The agentic system ASTRA (Agentic System for Tumor Resectability Assessment) was designed to process multimodal patient data, including magnetic resonance imaging (MRI) data, blood tests, and clinical notes, through a series of specialized agents: a data aggregator, a radiology report interpreter, and a multidisciplinary consensus simulator. The system then generates a structured MRI report and an assessment of radiologic and clinical unresectability.  

ASTRA was evaluated in a four-phase study. The first stage involved a retrospective evaluation of the system in 518 patients using a foundation model selection benchmark. This was followed by a reader study involving 104 patients that compared unaided vs ASTRA-assisted resectability assessment. The researchers then conducted prospective external validation in a cohort of 53 patients, comparing ASTRA’s performance with expert multidisciplinary team consensus. Finally, they conducted a silent trial in which ASTRA ran in real time in the background of a clinical workflow.  

Across the study stages, the researchers assessed the accuracy of resectability determinations, report structuring accuracy, inter-rater agreement, pre-multidisciplinary team report quality, and pre-multidisciplinary team briefing time.  

Evaluation Findings 

In the reader study, use of ASTRA improved the efficiency of report structuring for radiologists. Resectability assessment performance improved from 0.954 without assistance to 0.989 with ASTRA. Inter-rater agreement among radiologists also improved, with the Cohen’s kappa coefficient increasing from 0.887 to 0.974. In addition, the agentic system reduced radiologists’ assessment time by half.  

“A more critical part we were curious about is the impact on surgeons’ decision-making,” Dr. Huang noted. For surgeons, the agentic system improved resectability assessment performance from 0.827 without assistance to 0.952 with ASTRA. Accuracy in clinical resectability assessment also improved, from 0.788 to 0.904.  

Inter-rater agreement among surgeons increased as well, with the Cohen’s kappa coefficient rising from 0.654 to 0.933. Median assessment times were also reduced.  

In prospective validation, ASTRA achieved decision accuracy of 0.906 for radiologic unresectability assessment and 0.811 for clinical unresectability assessment. MRI reports were rated on completeness, correctness, necessity, and safety. The generated reports were considered equivalent in quality to routine multidisciplinary team consensus (4.26 ± 0.36 vs 4.39 ± 0.23).  

In the silent trial, the agentic system improved decision accuracy compared with unaided surgeons in determining radiologic and clinical unresectability. Report quality scores also increased from 2.79 to 4.26 with ASTRA assistance. Assessment time was nearly halved, from 518 seconds without ASTRA to 255 seconds with ASTRA. User acceptance survey results also improved.  

Dr. Huang noted that the outputs did not affect the clinical decision-making of the multidisciplinary team.  

Insights 

ESMO discussant Marco Lorenzi, PhD, a research scientist at the Inria Center of University Côte d’Azur, provided additional context on the use of agentic systems in oncology clinical decision-making.  

He cited a paper validating an autonomous AI agent capable of calling tools to perform different tasks. Tested on 20 realistic multimodal patient cases, the AI agent reached correct clinical conclusions in 91% of cases. Dr. Lorenzi explained that the current study goes beyond evaluating the reliability of tool calling, particularly with its prospective validation of the system.  

“What they’ve shown is actually that large language model–enhanced systems actually can outperform and lead to higher accuracy in this case,” Dr. Lorenzi commented. “This is a very encouraging and very promising validation of this kind of system in real life.” 

He cautioned, however, that such systems are complex and raise concerns about reliability, reproducibility, accountability, and security.  

Dr. Lorenzi noted that this is a rapidly evolving field and that further developments in the use of agentic AI systems for oncology clinical decision-making are expected in the near future.  

DISCLOSURES: Dr. Huang reported no personal or financial conflicts of interest. Dr. Lorenzi reported no personal or financial conflicts of interest. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.

KOL Commentary
Watch

Related Content