A cheap way to build a 3D chest CT diagnostic model is to borrow knowledge from a chest X-ray model. X-ray models are trained on hundreds of thousands of report-labelled images. CT scans with labels are scarce. Bootstrapping Chest CT Image Understanding by Distilling Knowledge from X-ray Expert Models (BIUD) does this. It "distills" a CT student from a 2D X-ray teacher, meaning the student is trained to reproduce the teacher's outputs. The two sides are linked by the similarity of their radiology reports, so no patient needs both scans. The catch is that X-ray models are known to cheat. Given chest X-ray pneumothorax, a baseline model reaches only 35.6% accuracy on its worst patient subgroup, because it keys on support devices such as chest tubes rather than on the collapsed lung (SpurCon). Standard multi-label training also lets one easy-to-see disease stand in for a harder co-occurring one (PAT). If the teacher cheats, the student may learn the same cheats. That could be one source of the over-sensitivity BIUD reports.
We propose to measure and remove this leakage. First we build a "leakage matrix" for the X-ray teacher, which records how much each disease output moves when a different finding is edited into or out of an image. The edits come from a generative model, and we accept only edits that pass a check confirming that only the intended finding changed. We then use the confirmed spurious directions to clean the teacher's outputs before distillation. The resulting CT student is evaluated on how often it predicts a disease because a correlated finding or device is present. The annotation cost is a small radiologist audit of generated edits, not dense CT labels.
Core Claim
Shortcuts learned by a 2D teacher transfer measurably into a 3D student under report-bridged distillation. Their fingerprint can be read off the teacher before any CT training, using counterfactual image pairs whose change has been verified. Projecting those fingerprints out of the teacher's per-disease outputs lowers the CT student's conditional false positives, meaning cases where disease B is predicted when only its usual companion A is present. It should do so without the accuracy loss seen when Unlearning Spurious Correlations in Chest X-ray Classification forced attention away from confounders (0.95 → 0.90).
The new element relative to BIUD is the audit-then-correct step. BIUD neither checks nor corrects what its teacher transfers. The new element relative to SpurCon is scope and pairing. SpurCon handles one binary confounder pair within a single modality and builds its pairs from patient IDs. We handle a full matrix of pathology-to-pathology and device-to-pathology leakage, build the pairs generatively, and evaluate in a different modality from the one audited.
Why verification is the crux
Generated counterfactuals only help if two things hold: the edit changed the intended finding, and the teacher's response is not driven by the generator's own pixel traces. The AI-generated-image detection literature treats the second problem directly. B-Free and Dual Data Alignment show that when a real image is compared with a synthetic one, differences in content and in generator artifacts get mixed together. Both papers fix this by passing the real image through the same generative pipeline, so the comparison happens at matched content. BPL adds that the comparison has to be made pairwise, not by scoring each image on its own.
We carry this over directly. Every edit $x_{A\to\neg A}$ (finding A removed or added) is compared against a null edit $\tilde{x}$: the same image reconstructed by the same generator with no change requested. The teacher's leakage from finding A onto disease head B is then
$$L_{AB} = \mathbb{E}\big[f_B(x_{A\to\neg A}) - f_B(\tilde{x})\big],$$
where $f_B$ is the teacher's score for disease B and the average runs over verified edits. Generator artifacts are present on both sides and cancel. Only the semantic change remains.
"Verified" means three things. First, an independent multi-label verifier agrees that A flipped and that no other label moved beyond a set tolerance. The verifier is trained on CheXpert with a different architecture from the teacher, which is trained on MIMIC-CXR. Second, the pixel difference between edit and null edit is concentrated where A plausibly appears. Third, radiologists audit a stratified sample of a few hundred pairs, which gives an estimate of the verifier's error rate. For that audit we reuse the odd-one-out protocol of Diffusion Visual Counterfactuals. There, ophthalmologists found good counterfactuals only 36.3% of the time against 33.3% chance, which gives us a concrete realism target. That work also supplies the technique: diffusion editing guided by a robust classifier. We apply it to multi-label chest X-rays, and we edit one finding while holding its companions fixed. This is exactly what the single-label "healing" in PathoGAN cannot do.
Research Program
Stage 1: build the edits. We train a report-conditioned latent diffusion editor on MIMIC-CXR. It covers BIUD's six CT-evaluable findings (nodule, opacity, effusion, calcification, consolidation, atelectasis) plus devices: chest tubes, lines and pacemakers. For each image and each target, it generates an edit and its null-edit control, and each pair passes through the verification gate.
Stage 2: estimate leakage. For the CheXzero-style teacher used by BIUD, we estimate the matrix $L$ from Stage 1. We also estimate spurious directions in the teacher's embedding space, $d_{A}$, defined as the mean embedding difference between verified edits and their null edits.
Stage 3: clean and distill. For each disease head B, we remove the components of the teacher's target that lie along the $d_A$ for which $|L_{AB}|$ exceeds a threshold. Distillation then proceeds exactly as in BIUD, including its RoCo contrastive loss, which treats patients with identical findings as positives rather than negatives.
Stage 4: diagnose. We test whether the teacher's leakage matrix predicts where the CT student fails. This is the scientifically interesting claim. If it holds, a cheap 2D audit becomes a pre-deployment check for 3D models.
On CT, the confounded groups come from report text: "effusion present, atelectasis absent", "tube present, no effusion or pneumothorax", and so on. A radiologist checks a subsample of these group assignments, because NLP labelers carry hedging noise.
Baselines
The naive reference is a CT model trained only on CT volume–report pairs, with no teacher at all. We use CT-CLIP from A foundation model utilizing chest CT volumes and radiology reports. It reports a zero-shot mean AUROC of 0.731 on CT-RATE internal validation, against 0.629 for a supervised CT-Net. This baseline cannot inherit X-ray shortcuts, so it sets both the floor for distillation's value and the reference level of confounding that CT data alone produces. We retrain it on the same CT volumes as our student.
The strongest existing method is BIUD itself, reproduced on ChestCT-16K and ChestCT-EXT. It is the only published X-ray-to-CT distillation for these six conditions. There is no reported AUROC for BIUD on Lacuna, so we report our own reproduction. We use the same backbone, epochs and hyperparameter search budget as our method, with only the teacher targets differing.
A third, strong competitor is BIUD with a SpurCon-debiased teacher. We train SpurCon's weighted contrastive adapter on the teacher using MIMIC patient IDs, which it needs and which MIMIC provides. SpurCon reports lifting CheXpert worst-group accuracy from 35.6% to 73.0%. We give it the same number of expert-labelled seed examples as our audit budget.
We also run FACT (FACT) on the student, as the representative of explicit comorbidity modelling. FACT reports +2.92% mAP on ODIR-5K and +2.25% on RSNA-IHD.
Four ablations isolate the key ingredients. The first drops the null-edit control and compares edits against originals. The second drops the verification gate and keeps all edits. The third projects out random directions of matched rank. The fourth projects out only device directions, to separate device leakage from pathology–pathology leakage.
Beating BIUD means two things together. The pooled conditional false positive ratio on CT confounded groups must fall by a margin outside the bootstrap confidence interval. Mean AUROC must fall by less than one point, and ideally rise.
Why not just…?
Why not just debias the teacher with SpurCon and skip the generator? This is the most serious threat, and it is run as a baseline. SpurCon's pairs require the same patient imaged with and without an artifact. That exists for chest tubes in MIMIC, but not for pathology–pathology pairs: no patient is re-imaged with effusion but without its accompanying atelectasis. SpurCon also handles one confounder at a time. If SpurCon-teacher distillation matches our CT conditional false positive rates on device groups, we will concede that axis and claim only the comorbidity pairs, where only generated edits can supply matched evidence.
Why not just train on CT alone and avoid inheriting anything? If the CT-only model matched distillation, the whole problem would disappear. CT-CLIP's gains suggest CT–report pretraining is strong, and BIUD claims its distillation adds on top. We report both. If CT-only already has the lowest conditional false positives and equal AUROC, that is itself a finding against X-ray distillation, and we would publish it as such.
Why not just strip device and companion-disease sentences from the report bridge, or mask devices? Report text only selects which teacher embeddings are used. The teacher's image features still encode a chest tube whether or not the text mentions it. SpurCon also found standard augmentation leaves worst-group accuracy at 0% on ISIC 2020, because the confounder sits inside the target region. We include report-sentence stripping as a cheap baseline. It should fail on groups where the confounder is visual but unreported.
Why not just model comorbidity explicitly, as FACT does? FACT builds its anchors from training co-occurrence rates. That encodes exactly the correlation we want the student not to lean on when disease A appears without disease B. We expect FACT to raise mAP but not lower conditional false positives. Its baseline run tests this.
Proposed Experiments
The first decisive check is whether leakage transfers at all. We compare the BIUD student with the CT-only model on the CT confounded groups. If the distilled student shows no excess conditional false positives on groups the teacher audit flags, the core hypothesis fails and we report it as a negative result.
The second check is predictive. Across label pairs, we compute the rank correlation between teacher leakage $L_{AB}$ measured on X-rays and the student's excess conditional false positive rate on CT. We consider a strong positive correlation, on held-out pairs, necessary for the audit to be useful.
The third check is the corrected student. We report mean AUROC and per-group AUROC on ChestCT-16K, ChestCT-EXT and CT-RATE, conditional false positive ratios as defined by PAT, and worst-group AUROC, all against the baselines above.
The fourth check tests the verification machinery. The null-edit and gate ablations should show that unverified or uncontrolled edits yield noisier leakage estimates, measured as lower split-half reliability of $L$, and worse CT correction. The radiologist audit should put the verifier's false-accept rate low enough that its errors cannot explain the observed effects.
References
- Bootstrapping Chest CT Image Understanding by Distilling Knowledge from X-Ray Expert Models
- SpurCon: Weighted Supervised Contrastive Learning for Mitigating Spurious Cues in Medical Imaging
- Counterfactual Reasoning for Multi-Label Image Classification via Patching-Based Training
- Unlearning Spurious Correlations in Chest X-Ray Classification
- A Bias-Free Training Paradigm for More General AI-generated Image Detection
- Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable
- BPL: Generalizable Deepfake Detection via Bias-only Pair-aware Learning
- CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison
- MIMIC-CXR: A large publicly available database of labeled chest radiographs
- Generating Realistic Counterfactuals for Retinal Fundus and OCT Images using Diffusion Models
- Pathology Segmentation Using Distributional Differences to Images of Healthy Origin
- A foundation model utilizing chest CT volumes and radiology reports for supervised-level zero-shot detection of abnormalities
- FACT: Fuzzy Alignment with Comorbidity Topology for Reliable Multi-Label Medical Image Diagnosis