Research line · ascendant
Medical: causal counterfactuals
Generate counterfactual images through an explicit causal model, not a conditional generator.
What defines membership
A medically meaningful 'what if this patient were older' requires causal machinery, and conditional generation alone does not provide it.
How the line developed
Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.
Before
Diff-SCM · 2022-02core
Brings diffusion into the causal counterfactual family, via forward diffusion for abduction and anti-causal gradient guidance for prediction. Experiments are MNIST and ImageNet only, so it is a methodological bridge rather than clinical evidence.
The idea
Deep SCM · 2020-06landmark
Normalizing flows and variational inference give the tractable invertibility that abduction, action and prediction counterfactuals need.
Evidence
PCGM · 2025-09core
An explicit causal graph plus voxel-level anatomical constraints for 3D brain MRI, whose counterfactual-derived measurements reproduce known disease effects from the neuroscience literature.
Improvement
Ribeiro et al. · 2023-06landmark
The direct successor to the origin paper, and the reference for judging counterfactuals by causal axioms rather than distribution distance. It validates two of them, effectiveness and composition, on Morpho-MNIST, UK Biobank brain MRI and MIMIC-CXR. Later papers score against four; this one did not establish all four.
Causal-Adapter · 2025-09strong-followup
Puts structural-causal-model interventions inside a frozen text-to-image diffusion model, on ADNI brain MRI, scored on effectiveness, composition, realism and minimality.
Limitation
Singla et al. · 2022-12core
A GAN counterfactual explainer with no causal model, validated by a reader study with 12 diagnostic radiology residents: counterfactual explanation was the only style that significantly improved their understanding of the classifier. Note the paper does not claim causal machinery is unnecessary. Its own conclusion says counterfactual reasoning is incomplete without a causal structure, and leaves that as future work. It challenges the line on evidence while conceding its premise.
Shortcut counterfactuals · 2023-12strong-followup
Classifier-guided diffusion, no causal model, used to find and quantify shortcut features like pacemakers across two chest X-ray datasets.
StylEx (medical) · 2024-04strong-followup
States outright that it is not designed to infer causality, then produces clinically useful counterfactual attribute discovery across eight tasks, expert-panel validated.
DiffChest · 2024-09core
515,704 radiographs from 194,956 patients, no causal graph, latents perturbed by a linear-classifier gradient. Fleiss' kappa at or above 0.8 for confounder detection, plus downstream diagnostic benchmarking against CheXzero. The largest-scale evidence here comes from work that builds no causal machinery.
What it gets right
- Answers a question conditional generators cannot even pose
- Evaluated against causal axioms rather than distribution distance
- Directly relevant to fairness and treatment-effect questions
Where it is weak
- Requires an explicit causal graph, which is hard to specify for imaging
- Sample quality lags behind mainstream diffusion
- The strongest clinical evidence in this area comes from work that uses no causal machinery at all: a 12-resident reader study, a 194,956-patient cohort, and an expert-panel study. The causal papers are evaluated on axioms, the non-causal ones on clinicians. That asymmetry is the line's central unresolved problem. It is a gap in what the causal side has demonstrated rather than a refutation of its premise, and the strongest challenger explicitly agrees that causal structure is ultimately needed