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

Amy Rafferty, Ajitha Rajan, Rishi Ramaesh

University of Edinburgh, UK

NHS Lothian, UK

Motivation

• Deep learning models for chest X-ray pathology classification have achieved high performance
• Complex black boxes – limits trust for clinical deployment
• Concept Bottleneck Models improve interpretability by introducing a human-readable middle step
• E.g., predicting clinical concepts before diagnoses

​

Motivation

• Standard CBMs treat concepts as independent predictive features
• Model learns which concepts correlate with which diagnoses
• Concept -> Pathology

​

• Causal CBMs instead learn relationships between concepts and pathologies
• Causal graph
• Direction and structure learned from data (typically Concept -> Pathology)
• Can lead to models learning spurious or clinically implausible relationships (e.g., Pneumothorax and chest tubes)

​

• Each image-report pair in the MIMIC-CXR dataset is converted into a binary concept vector indicating the presence/absence of each concept
• Radiologist-defined concept space (publicly available) – concepts correspond to key phrases extracted from the report paired with each image
• 11 concepts: Unremarkable, Mass, Nodular Disease, Irregular Hilum/Mediastinum, Pneumonitis, Consolidation, Opacities, Infection, Pleural Disease, Enlarged Heart, Absent Lung Markings

Concept Vectors

• Six target pathologies: No Relevant Finding, Cardiomegaly, Suspicious Malignancy, Pneumonia, Pneumothorax, Pleural Effusion
• Radiologist-defined associations: Strong (2), Weak (1), None (-)

Pathology-Concept Associations

XpertCausal

• Radiologist-constrained noisy-OR causal graph
• Multiple pathologies can independently cause the same finding (concept)
• Radiologist-defined associations determine which edges are allowed in the causal graph
• Allowed edge strengths initialized using association matrix, refined during training

XpertCausal

• Bayesian Inference for final diagnosis prediction
• Model is trained in true causal direction (P(C|Y)), but final diagnosis requires P(Y|C)
• Real diagnostic reasoning: pathologies cause concepts, which are then treated as evidence

XpertCausal

• Provides explanations of its predictions based on causal contributions
• For the predicted pathology:
• Measure how much each concept contributed to its probability
• Returns the two concepts with highest log-likelihood contributions

Evaluation: Methods

• XpertCausal achieved the highest pathology prediction performance, and lowest Expected Calibration Error (ECE)
• Causal structure improves performance even without radiologist constraints, but incorporating the constraints gives larger gains

​

​

​

​

​

​

• Explanation Quality: XpertCausal top-K concept explanations consistently contained the highest proportion of ground truth clinical concepts at low values of K
• Explanations should be concise and relevant – performance at low K is critical

Evaluation: Results

• XpertCausal’s causal graph more clearly separates the pathology-concept associations defined by the radiologist as Strong (2), Weak (1) and None (-)
• The causal ablation model (graph learned from data only) assigns weights to clinically implausible associations
• E.g., Pneumothorax -> Unremarkable (0.4630), Suspicious Malignancy -> Unremarkable (0.3374)

Evaluation: Results

Conclusion

• XpertCausal combines traditional CBMs with probabilistic causal disease modelling
• Radiologist-defined constraints prevent implausible concept-pathology relationships
• XpertCausal improves pathology prediction performance, model calibration, explanation relevance and alignment with real-world clinical domain knowledge
• Clinically motivated causal structure

Ongoing Work

• Application of XpertCausal to high-quality NHS data through Public Health Scotland
• Aim to assess performance on real data (~100k chest X-rays, CT scans and accompanying radiology reports)
• Incorporate a formal radiologist evaluation of explanation quality
• Extend concept space and graph design to include multiple radiologists

Thank you!