Before cancer surgery, doctors estimate how far the tumor has spread from CT or MRI scans. This is the clinical stage. After surgery, a pathologist examines the removed tissue under a microscope and assigns the pathological stage. Models that predict stage from scans are trained to match that pathology label. The label, however, is only as good as the surgery that produced it. If the surgeon removes and examines too few lymph nodes, a patient whose cancer has spread to the nodes can be recorded as node-negative. In esophageal cancer, fewer than 18 examined nodes measurably shifts patients into falsely reassuring categories (Prognostic impact of examined lymph-node count for patients with esophageal cancer). A model trained on these labels learns to under-stage. Under-staging is the more dangerous error, because it leads to withholding chemotherapy.
We propose to train a model that reads preoperative scans and writes a structured staging report. Its training signal is broken into separate checklist items, a design known as reward decomposition. A grader that has read the postoperative pathology report scores each item, and each item's weight depends on how much the pathology can actually confirm that claim for this patient. Tumor depth measured on the resected primary is trusted fully. A "node-negative" pathology result from an eight-node dissection is trusted only partially, by an amount set by a simple sampling model. Statements that pathology never checks, such as distant metastasis in most cases, earn no reward at all. The two-sentence version: pathology is an answer key with holes of known size, so we score the model only against the parts of the key that are filled in, and we weight each part by how completely it is filled.
Core Claim
Label corruption in staging is not uniform. It differs across the parts of a stage and it is predictable from recorded surgical facts. Published staging models, whether multi-task networks, detectors, or survival forests, treat the pathology label as a single trusted target. We claim that a per-item reward, gated by a sampling-adequacy model, will reduce preoperative nodal under-staging without raising over-staging, compared with the same model trained on raw pathology labels.
We also make a second, weaker claim that we treat as something to test. Reinforcement learning (RL) against a grader that reads the report matters only when the pathology evidence is free text that varies in form from case to case. If the pathology facts can be cleanly pulled out into fixed labels, an ordinary supervised loss with the same adequacy weights may do just as well. We design the experiments so that this outcome is visible.
Where the Idea Comes From
Three lines of work on reward decomposition supply the method. Checklists Are Better Than Reward Models For Aligning Language Models replaces a single learned reward score with instruction-specific yes/no checklist items. 2D-DPO scores each sentence of a response on several quality criteria separately. Rubric-Grounded Reinforcement Learning introduces "information asymmetry": the policy sees only the question, while a frozen grader sees a hidden reference passage and a weighted rubric. We reuse that asymmetry directly. The preoperative model sees the scan, and the grader sees the pathology report and the examined-node count.
All three assume that the hidden reference is correct. Our addition is a reference whose reliability varies by item and by patient, with that reliability computed from surgical records rather than assumed. InfiMed-ORBIT is the closest medical precedent. It builds case-specific clinical rubrics with separate positive items and negative "red flag" items, and trains with GRPO (a policy-gradient method that compares several sampled answers per input). It works only on text dialogue, and it constrains our design. It reports that heavy penalties on safety omissions push policies toward hedging and disclaimers (ORBIT, second entry). A staging model could hedge in the same way, by calling every node indeterminate. For that reason we measure the indeterminate rate explicitly and cap any reward for abstaining.
The staging literature supplies the motivation and the rubric content. Node-RADS on preoperative CT scores nodes on size, shape, border and internal texture. It separates prognosis far better than binary node status (5-year progression-free survival AUC 0.850 versus 0.643). The same paper notes that only 3% of nodes received pathological confirmation, which is the verification gap in its starkest form. Ratio-based node metrics such as the positive-node ratio and log odds of negative nodes (pancreatic nodal staging; oral cavity LNR) correct for the number of nodes retrieved, but only after surgery and only for predicting prognosis. We move that correction into the training signal of a model that runs before surgery.
Research Program
The policy is a vision-language model that takes preoperative CT and writes a fixed-schema report. The report contains a Node-RADS attribute tuple for each detected node, a verdict for each nodal station, a cN category, a cT depth estimate, and a flag for suspected distant metastasis. The grader is a frozen language model. It receives the report, the pathology report, the examined-node count, and a rubric generated for the case. This follows ORBIT's retrieval-based rubric construction, restricted to staging items.
Each item $i$ receives a score $s_i \in \{0,1\}$ from the grader and a trust weight $w_i \in [0,1]$. The case reward is
$$R = \frac{\sum_i w_i\, \lambda_i\, s_i}{\sum_i w_i\, \lambda_i},$$
Here $\lambda_i$ is a fixed clinical cost that is larger for items whose failure means under-staging. Items with $w_i = 0$ drop out of the reward entirely.
The main new element is the trust weight for a "pathology node-negative" item. Suppose the regional basin holds $N$ nodes, $k$ of them positive, and the surgeon examined $n$ nodes chosen at random. The chance that pathology sees no positive node is $\binom{N-k}{n}/\binom{N}{n}$. This is the hypergeometric probability of drawing only negative nodes. Combining it with a prior on $k$ estimated from adequately dissected cases gives the posterior probability that the patient is truly node-negative given pN0 and $n$. That posterior is used as $w_i$. In practice, when $n$ is small the model is not heavily penalized for flagging a suspicious node that pathology called negative. It is still fully penalized for calling a node negative when pathology found it positive, because a positive finding is never a sampling artifact.
We then train with GRPO, with a KL penalty (a term that keeps the policy close to its starting model), and run ablations that isolate the adequacy gating.
Baselines
The naive reference point is CoSAM (Meply), a detector followed by a prompt-based segmenter for metastatic perirectal nodes. Its patient-level N call is made by thresholding its detections. The paper reports Dice 74.12%, IoU 58.59% and detection AP50 0.875 on Meply. It is the strongest published imaging method in the file for this task. We retrain it on our splits with its published recipe. It counts as beaten only if we lower patient-level under-staging at matched over-staging without losing more than one Dice point.
The strongest competitor on our actual question is a supervised multi-task model. It follows the design of the cervical cancer multi-task ANN, which reached 76% LNM accuracy, but uses our CT backbone. It has heads for the Node-RADS attributes, station, cN, cT and M. Missing labels are masked. Each loss term is multiplied by the same trust weights $w_i$ and costs $\lambda_i$ that our reward uses. This baseline is identical to our method except for the RL-with-grader step, so it isolates exactly that ingredient. It receives the same backbone, data and number of hyperparameter trials as our method, with GPU-hours matched to include the grader's inference cost.
We also run a cost-sensitive version that uses unweighted pathology labels, the same model and objective as our method with all trust weights set to one, and an RL variant that uses a single scalar "stage correct" reward. For tabular context we include a random survival forest on clinical features and examined-node count. The esophageal study reports AUC 79.3 on external testing for this approach. No reported result was found on Lacuna for any vision-language model producing structured staging reports.
Why not just…?
Why not just train a supervised multi-task network with auxiliary Node-RADS heads? This is the real threat. If the pathology facts can be extracted cleanly into fixed labels, masked supervised losses carrying the same trust weights may match RL. The trust-weighted supervised baseline above tests this directly. If it ties, we will report that the gain comes from adequacy gating and not from reward decomposition. We expect RL to help only in the portion of cases where the pathology reports are free text with station information that varies in form, since labels there cannot be mapped reliably to fixed heads.
Why not just use a cost-sensitive loss that penalizes false negatives more? A global weight on false negatives cannot tell a truly negative patient apart from a patient labeled negative after an eight-node dissection. It therefore pushes all node calls toward positive and raises over-staging, which is the error that wrongly denies patients curative surgery (stage migration). The cost-only ablation measures this trade-off. Our claim fails if that ablation reaches the same operating point.
Why not just use lymph-node ratio nomograms or survival forests? These models need the examined-node count and positive-node count, which exist only after surgery, so they cannot guide the preoperative decision itself. We include the survival forest to show how much prognostic information the tabular data already holds, not as a competitor for preoperative staging.
Proposed Experiments
The decisive controlled check is semi-synthetic and uses Meply, whose metastatic nodes carry expert pixel-level annotation. We treat those annotations as the truth. We then simulate surgery by drawing $n$ examined nodes per patient from 4 to 25. The resulting pN label is positive only if at least one annotated metastatic node is drawn. We train every method on these corrupted labels and evaluate against the full annotations.
Our prediction is that the under-staging gap between trust-weighted methods and unweighted methods widens as $n$ falls, and that it disappears near complete sampling. If the gap does not track $n$, the central claim is wrong.
We run the same simulation with corruption that ignores sampling adequacy. This checks that the gain comes from correctly modeling how labels go missing and not from generic label smoothing.
The second experiment tests clinical validity on a CT cohort paired with pathology reports that record examined-node counts. The file lists no public dataset of this kind, so the experiment requires an institutional collection similar to the multicenter pRCC cohort. There we report under-staging at fixed over-staging, calibration of cN, the indeterminate-output rate as a check on hedging, and the ability of predicted cN to separate progression-free survival. The survival analysis matters because it evaluates the model against outcomes rather than against the flawed pathology label.
References
- Prognostic impact of examined lymph-node count for patients with esophageal cancer: development and validation prediction model
- Checklists Are Better Than Reward Models For Aligning Language Models
- 2D-DPO: Scaling Direct Preference Optimization with 2-Dimensional Supervision
- Rubric-Grounded Reinforcement Learning:Structured Judge Rewards for Generalizable Reasoning in Language Models
- InfiMed-ORBIT: Aligning LLMs on Open-Ended Complex Tasks via Rubric-Based Incremental Training
- Node-RADS category on preoperative CT predicts prognosis in patients with papillary renal cell carcinoma
- Predictive value of current nodal staging systems and development of machine learning nomogram for resectable pancreatic head cancer: a population-based study and multicenter validation
- Prognostic value of positive lymph node ratio in oral cavity squamous cell carcinoma after neoadjuvant treatment: a retrospective real-world study
- Meply: A Large-scale Dataset and Baseline Evaluations for Metastatic Perirectal Lymph Node Detection and Segmentation
- A multi-Task Learning based applicable AI model simultaneously predicts stage, histology, grade and LNM for cervical cancer before surgery
- A Case of Benign Metastasizing Leiomyoma of the Lung Complicated with Primary Lung Cancer