Lung cancer clinical decision support demands precise reasoning across complex, multi-stage oncological workflows. Existing multimodal large language models (MLLMs) fail to handle guideline-constrained staging and treatment reasoning. We formalize three oncological precision treatment (OPT) tasks for lung cancer, spanning TNM staging, treatment recommendation, and end-to-end clinical decision support. We introduce LungCURE, the first standardized multimodal benchmark built from 1,000 real-world, clinician-labeled cases across more than 10 hospitals. We further propose LCAgent, a multi-agent framework that ensures guideline-compliant lung cancer clinical decision-making by suppressing cascading reasoning errors across the clinical pathway. Experiments reveal large differences across various large language models (LLMs) in their capabilities for complex medical reasoning, when given precise treatment requirements. We further verify that LCAgent, as a simple yet effective plugin, enhances the reasoning performance of LLMs in real-world medical scenarios.
Alex: Welcome to another episode of ResearchPod.
Sam: Today, we're looking at a paper on using AI to help with lung cancer decisions. It introduces LungCURE, a benchmark with 1,000 real patient cases from over 10 hospitals, labeled by doctors. This tests how well AI models handle lung cancer workflows, like figuring out the cancer stage and suggesting treatments that follow strict medical rules.
Alex: So this is about AI making mistakes in something as serious as cancer staging?
Sam: Yes. Doctors classify lung cancer using TNM staging. T looks at the tumor's size and how far it's grown into nearby tissues. N checks if cancer has spread to lymph nodes near the lungs. M sees if it has reached distant body parts like the liver or bones. Getting this wrong leads to bad treatment choices that can shorten lives, as the paper shows with survival graphs.
Alex: And these AI models read patient reports, images, and lab results to do this staging and recommend therapies?
Sam: They're multimodal large language models—computer systems trained on huge amounts of text and images to understand mixed data, like a doctor's notes plus scan descriptions. In a rural clinic, you might upload CT scans and reports for quick advice. But the paper finds these models often hallucinate—they invent wrong details, like confusing node spread with tumor size. That leads to picking the wrong chemo instead of targeted drugs for gene mutations.
Alex: So one staging slip-up messes up the whole treatment plan?
Sam: Exactly. Errors propagate because everything's handled in one big reasoning step, mixing up evidence across T, N, and M. The paper proposes a multi-agent setup to fix that. First, the benchmark reveals big gaps in current models' performance.
Alex: How does this multi-agent setup stop those errors from snowballing?
Sam: It breaks the big decision into smaller, specialized steps—like dividing a team project so each person handles one clear part without overlap. Instead of one AI juggling everything, it splits into focused agents. One checks only tumor size and local growth. Another looks at nearby lymph nodes. A third checks distant spread. These outputs feed into a final step that combines them using official medical logic. This prevents mistakes in one area from infecting the next, like containing a spill before it spreads.
Alex: So it's like separate checklists for each body part, to avoid confusing a spot on the lung with one far away?
Sam: Precisely. The paper calls this anatomical dimension isolation. Each agent gets prompts limited to its own evidence, so the tumor agent ignores node or distant clues. They pull out key patient details like gene mutations or scan findings, and route them only to the right agents with matching rules. With one top model, overall decision precision rose about 30% compared to just prompting the AI directly—mainly because treatment suggestions build on cleaner staging.
Alex: Before those T, N, and M agents get their checklists, how do they start with messy reports like scans and notes—and turn them into something usable?
Sam: It begins with cleaning and organizing raw reports—like sorting puzzle pieces by color before assembling, so no wrong fits later. For a CT scan, it splits into what the doctor saw and a summary impression, keeping exact sizes and uncertain words like "might be spread." It breaks combined phrases, such as nodes in multiple spots, into separate items. The paper stresses preserving doubts exactly as written.
Alex: So that creates a neutral list each agent can use, without guessing too early?
Sam: Yes. A filter grabs only tumor clues—like size or pushing into chest wall—for the T agent, ignoring the rest. Picture handing a builder just the foundation blueprint. The T agent checks rules step by step: size brackets like under 3 cm, then invasions, picking the highest match. A dispatch step sorts node clues into regional or distant pools, keeping them pure. The N agent scans for confirmed spread: nearby nodes make N1, deeper chest N2. For M, it checks chest spread first, then distant organs. All output structured facts for the final combo using AJCC tables—like a calculator for scores. It flags uncertainties too, alerting doctors.
Alex: So narrow views plus rule checklists stop one bad read from tainting others.
Sam: That's the point. This cuts interference, like a tumor hint fooling node judgment, leading to cleaner stages that feed accurate treatments.
Alex: Once staging is set, how does it handle picking treatments without drifting off guidelines?
Sam: A step pulls key details—like cancer cell type, fitness, gene changes—from reports and staging into an organized list. It's like filling a standard form with exact categories. This list routes to the right specialist agent—for post-surgery cases, pre-surgery planning, or advanced stages by gene type. Each gets only matching guidelines, like a narrow playbook. A post-surgery agent sticks to rules for adding chemo after tumor removal. An advanced agent uses PD-L1 levels for immunotherapy versus chemo. Driver-positive ones match genes to targeted therapies. This keeps outputs tied to evidence from proven trials.
Alex: Are there spots where it still leans on the base model's limits?
Sam: Yes. It depends on the model's staging accuracy, around 60 to 75 percent—stronger with text than images. Gains are meaningful but tied to that base. It hasn't been checked against guidelines after 2025, so tweaks might be needed. Still, the paper suggests it's a practical bridge, turning shaky outputs into guideline-faithful steps.
Alex: Yeah... a grounded way to harness these models without overpromising. Thanks for breaking it down, Sam—that clarifies the full picture.
Sam: My pleasure, Alex. Thanks for listening to ResearchPod.