Research map / Probabilistic and scientific modeling / Treatment effects
Learning Causal Alignment for Reliable Disease Diagnosis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a causal alignment loss that uses counterfactual generation and implicit-function-theorem gradients (with conjugate gradient) to align a medical classifier's causal decision factors with radiologist annotations, with a hierarchical attribute-based variant.
Problem
Medical image classifiers learn associational rather than causal alignment with radiologists' decision processes, so they rely on pseudo-correlated shortcut features that fail to transfer.
Approach
Identify causal factors of each decision via counterfactual image generation, then penalize counterfactual modifications falling outside radiologist-annotated regions; differentiate through the implicit counterfactual generator using IFT and conjugate gradient; extend to a hierarchical image→attribute→label pipeline with causal attribution (CCCE).
Key results
- On LIDC-IDRI, the method reaches CAM precision 0.751 and accuracy 0.722, versus best baseline CAM precision 0.074 and accuracy 0.656.
- On CBIS-DDSM, CAM precision 0.805 and accuracy 0.656, versus best baseline CAM precision 0.232 and accuracy 0.592.
- Ablation: the causal alignment loss raises LIDC CAM precision from 0.057 to 0.587; hierarchical alignment adds further gains to 0.751.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add causal alignment loss (ablation) | Causal alignment method without the alignment loss · LIDC-IDRI | CAM precision: 0.057 (0.07) -> 0.587 (0.08) | Helped |
| add causal alignment loss (ablation) | Causal alignment method without the alignment loss · CBIS-DDSM | CAM precision: 0.143 (0.20) -> 0.621 (0.03) | Helped |
| add causal alignment loss (ablation) | Causal alignment method without the alignment loss · LIDC-IDRI | classification accuracy: 0.535 (0.08) -> 0.701 (0.02) | Helped |
| add causal alignment loss (ablation) | Causal alignment method without the alignment loss · CBIS-DDSM | classification accuracy: 0.592 (0.00) -> 0.633 (0.03) | Helped |
| add hierarchical alignment (ablation) | Causal alignment loss without hierarchical alignment · LIDC-IDRI | CAM precision: 0.587 (0.08) -> 0.751 (0.03) | Helped |
| add hierarchical alignment (ablation) | Causal alignment loss without hierarchical alignment · CBIS-DDSM | CAM precision: 0.621 (0.03) -> 0.805 (0.06) | Helped |
| add hierarchical alignment (ablation) | Causal alignment loss without hierarchical alignment · LIDC-IDRI | classification accuracy: 0.701 (0.02) -> 0.722 (0.00) | Helped |
| add hierarchical alignment (ablation) | Causal alignment loss without hierarchical alignment · CBIS-DDSM | classification accuracy: 0.633 (0.03) -> 0.656 (0.00) | Helped |
Surprising findings
- The Oracle classifier, restricted to radiologist-annotated areas, reaches only 72-79% accuracy, far below the >99% claimed in some prior work, attributed to excluding hard samples and different splits.
Limitations
- Causal alignment loss optimization is computationally expensive due to implicit Jacobian estimation.
- Evaluated only on two medical classification datasets with small sample sizes (under 1,000 images).
Experimental setup
Datasets: LIDC-IDRI CBIS-DDSM BraTS Breast Ultrasound Images (Aryashah2k)
Benchmarks: LIDC-IDRI CBIS-DDSM BraTS Breast Ultrasound
Metrics: CAM precision relative to annotated areas classification accuracy
Baselines: input gradient regularization (Ross et al.) ICNN BagNet multi-task learning with explanations counterfactual data augmentation Oracle classifier
Builds on: implicit function theorem conjugate gradient structural counterfactuals (probability of causation) causal attribution (CCCE)
Closest papers by meaning
- Conformal Alignment: Knowing When to Trust Foundation Models with Guarantees (NeurIPS 2024 · similarity 0.85)
- Learning Counterfactual Outcomes Under Rank Preservation (NeurIPS 2025 · similarity 0.83)
- Counterfactual Generative Modeling with Variational Causal Inference (ICLR 2025 · similarity 0.83)
- S2D-Align: Shallow-to-Deep Auxiliary Learning for Anatomically-Grounded Radiology Report Generation (AAAI 2026 · similarity 0.83)
- Measure-Theoretic Anti-Causal Representation Learning (NeurIPS 2025 · similarity 0.83)
- Automatic Radiology Reports Generation via Memory Alignment Network (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.