💻 computer science

Radiologist-Guided Causal Concept Bottleneck Models for Chest X-Ray Interpretation

The paper introduces XpertCausal, a radiologist-guided causal Concept Bottleneck Model that leverages a probabilistic noisy-OR framework and expert-defined constraints to better align chest X-ray interpretation with clinical reasoning, resulting in improved accuracy, calibration, and interpretability compared to existing baselines.

Original authors: Amy Rafferty, Rishi Ramaesh, Ajitha Rajan

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Amy Rafferty, Rishi Ramaesh, Ajitha Rajan

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to read chest X-rays, just like a doctor does. The goal is for the computer to not only say, "This patient has pneumonia," but also to explain why it thinks that, using language a human can understand.

This paper introduces a new system called XpertCausal. To understand how it works, let's use a few analogies.

The Problem: The "Black Box" and the "Wrong Direction"

Most current AI models are like black boxes. You feed an X-ray in, and a diagnosis pops out. You have no idea what the computer was looking at to make that decision.

To fix this, researchers created "Concept Bottleneck Models" (CBMs). Think of these as a middleman. Instead of jumping straight from X-ray to diagnosis, the AI first says, "I see a white spot," and "I see an enlarged heart." Then, it uses those observations to guess the disease.

However, the paper argues that most of these middlemen are thinking in the wrong direction.

The Old Way (Discriminative): The AI thinks, "If I see a white spot, that means pneumonia." It treats the spot as a clue to find the disease.
The Real World (Generative): In reality, the disease comes first. Pneumonia causes the white spot. The spot is the result, not the cause.

The paper says that by ignoring this "cause-and-effect" flow, AI models can get confused. They might learn fake connections (like thinking a white spot always means pneumonia, even when it doesn't) because they are just looking for patterns in the data without understanding the biology.

The Solution: XpertCausal

The authors built XpertCausal, which flips the script. Instead of just looking for clues, it tries to simulate how a disease actually creates an X-ray image.

Here is the analogy:
Imagine a recipe book (the disease) that tells you exactly what ingredients (the X-ray findings) will appear in the final dish.

Pneumonia is the recipe.
White spots and fluid are the ingredients that must appear if you follow that recipe.

XpertCausal uses a Radiologist's Guidebook (a matrix created by a human expert) to define these recipes. The guidebook says: "If the patient has Pneumonia, there is a strong chance of seeing a white spot. If they have a broken rib, there is a weak chance of seeing a specific shadow."

The AI then works backward. It looks at the X-ray, identifies the ingredients (the spots and shadows), and asks: "Based on my recipe book, which disease is most likely to have created this specific combination of ingredients?"

How It Uses the "Guidebook"

The key innovation is that the AI doesn't just guess the connections; it is constrained by a human expert.

Without the Guidebook (The "Learned" Model): If you let the AI figure out the connections on its own, it might get creative in a bad way. It might decide that "Pneumonia" causes "No findings at all" just because it saw that weird pattern once in the data.
With the Guidebook (XpertCausal): The human expert tells the AI, "No, Pneumonia never means 'No findings'." This stops the AI from learning silly or dangerous rules.

The Results: Why It Matters

The researchers tested this system on a huge database of real chest X-rays (MIMIC-CXR) and compared it to two other versions:

The Standard AI: Just looks for patterns (Black Box).
The "Middleman" AI (XpertXAI): Uses human concepts but doesn't understand cause-and-effect.
The "Unconstrained" AI: Uses the cause-and-effect logic but learns the rules from scratch without a human guide.

XpertCausal won in three main ways:

It was more accurate: It diagnosed diseases better than the others.
It was more honest: It was better at knowing when it was unsure (calibration).
It gave better explanations: When asked "Why did you say Pneumonia?", it pointed to the most relevant spots on the X-ray first. The other models often pointed to less important things or got confused.

The Bottom Line

The paper claims that by teaching the AI to think like a doctor (understanding that Disease $\rightarrow$ Findings) and giving it a human-written rulebook to prevent it from learning nonsense, we get a system that is not only smarter but also easier to trust. It doesn't just guess; it reasons through the "recipe" of the disease, making its decisions clearer and more reliable for medical use.

Technical Summary: Radiologist-Guided Causal Concept Bottleneck Models for Chest X-Ray Interpretation

Problem Statement
Concept Bottleneck Models (CBMs) aim to enhance interpretability in medical imaging by predicting intermediate, human-understandable concepts before final diagnoses. However, existing CBMs in this domain face three primary limitations:

Discriminative Bias: Most treat concepts as independent predictive features (Concept $\rightarrow$ Pathology), ignoring the underlying clinical generative process where diseases cause observable findings (Pathology $\rightarrow$ Concept).
Lack of Clinical Grounding: Many approaches rely on weakly supervised or latent concepts that may not align with meaningful diagnostic reasoning or stable semantic definitions.
Spurious Associations: Without explicit structural constraints, models trained on noisy medical datasets may learn implausible correlations between concepts and pathologies that do not reflect real-world diagnostic logic.

Methodology: XpertCausal
The authors propose XpertCausal, a radiologist-guided causal CBM for chest X-ray (CXR) interpretation. The architecture follows a two-stage pipeline:

Concept Prediction: An InceptionV3 backbone (pretrained on ImageNet) maps input CXRs to a vector of 11 clinically curated concepts (e.g., "Opacities," "Enlarged Heart") derived from radiology reports via a rule-based extraction pipeline.
Causal Pathology Inference: Unlike standard CBMs, XpertCausal models the relationship in the generative direction (Pathology $\rightarrow$ $\to$ Concept) using a noisy-OR probabilistic framework.
- Generative Model: The probability of a concept $c_k$ being present given a set of pathologies $y$ is modeled as:
  $P(c_k = 1 | y) = 1 - (1 - \lambda_k) \prod_{j: y_j=1} (1 - \eta_{jk})$
  where $\eta_{jk}$ is the association strength between pathology $j$ and concept $k$ , and $\lambda_k$ is a leak term.
- Expert Constraints: The model structure is constrained by a radiologist-curated association matrix. This matrix defines relationships as Strong (2), Weak (1), or None (-), ensuring that only clinically plausible pathways exist in the causal graph.
- Bayesian Inference: At inference, the generative model is inverted to estimate pathology probabilities given the predicted concepts:
  $P(y | c) = \frac{P(c | y) P(y)}{\sum_{y'} P(c | y') P(y')}$
- Explanation Generation: Explanations are derived by ranking concepts based on their causal contribution (log-likelihood ratios) to the predicted pathologies, rather than just prediction confidence.

Key Contributions

Novel Architecture: Introduction of XpertCausal, which combines interpretable concept bottlenecks with probabilistic causal disease modeling using a noisy-OR framework and Bayesian inversion.
Expert-Guided Structure: Integration of radiologist-defined concept-pathology associations as structural constraints within the causal graph. This enforces clinically grounded reasoning and prevents the learning of implausible dependencies.
Comprehensive Evaluation: A rigorous evaluation on the MIMIC-CXR dataset comparing XpertCausal against a non-causal expert-driven baseline (XpertXAI) and a causal ablation model with unconstrained learned associations (Causal (Learned)).

Results
Evaluated on the MIMIC-CXR dataset (44,038 CXR-report pairs), XpertCausal demonstrated superior performance across multiple metrics:

Pathology Prediction: XpertCausal achieved the highest macro-averaged AUROC (0.7968) and F1 Score (0.7109), outperforming the non-causal XpertXAI (AUROC 0.7361) and the unconstrained Causal (Learned) model (AUROC 0.7631). It also achieved the best calibration (Expected Calibration Error of 0.0079).
Explanation Quality: XpertCausal captured the highest proportion of ground-truth concepts at low $K$ (top 1–2 concepts), indicating it more reliably ranks diagnostically relevant findings as the most important.
Graph Faithfulness: Analysis of learned edge weights showed that XpertCausal maintained strong separation between "Strong," "Weak," and "None" categories defined by experts. In contrast, the unconstrained Causal (Learned) model assigned non-zero weights to "None" associations (e.g., linking "Pneumothorax" to "Unremarkable"), reflecting spurious correlations rather than clinical logic.

Significance and Claims
The paper claims that incorporating clinically motivated causal structure and expert domain knowledge into CBMs leads to models that are not only more accurate and better calibrated but also more aligned with human diagnostic reasoning. By explicitly modeling the generative process (Pathology $\rightarrow$ Concept) and inverting it via Bayesian inference, XpertCausal provides a framework where explanations are tied to the causal mechanism of disease manifestation rather than mere statistical correlation. The authors assert that causal structure alone is insufficient in noisy medical datasets; without expert-defined constraints, learned graphs risk capturing dataset-specific artifacts rather than true diagnostic pathways.

Limitations
The authors acknowledge that the current study is limited to a specific concept space (11 concepts) and six target labels curated by a single radiologist. Future work is proposed to expand the concept and pathology spaces, incorporate additional expert relationships, and validate generalization across external datasets and imaging settings.