Concept Bottleneck Model

A CBM decomposes inference into two distinct, auditable stages. Stage 1 maps raw input (e.g., a chest X-ray) to a set of human-specified concept scores, such as 'bone lesion present,' 'lung opacity,' or 'pleural effusion.' Stage 2 uses only these concept scores—not the raw pixels—to make the final diagnostic prediction. This bottleneck forces all reasoning through the concept layer, making the decision path fully transparent. A radiologist can inspect exactly which concepts activated and to what degree, enabling direct clinical validation of the model's reasoning before accepting its output.

A defining capability of CBMs is that clinicians can directly edit concept predictions at test time to correct model errors and immediately improve final predictions. If the model incorrectly predicts a high score for 'cardiomegaly' when the heart is normal, a radiologist can override that concept value to 'absent.' The final diagnosis is then recalculated using the corrected concept set, without retraining. This clinician-in-the-loop paradigm transforms the AI from a black-box oracle into a collaborative tool where human expertise can steer and correct model behavior in real time, building trust and improving accuracy.

Training a CBM requires concept-level supervision—each training image must be labeled not just with the final diagnosis but also with ground-truth values for each intermediate concept. This creates a higher annotation burden compared to standard end-to-end models. Strategies to mitigate this include:

• Using expert-defined concept vocabularies from medical ontologies like RadLex or SNOMED CT
• Leveraging synthetic concept labels generated by existing classifiers or large vision-language models
• Employing incomplete concept supervision where only a subset of concepts are labeled per sample
• Using concept-level data augmentation to expand limited annotations

The second stage of a CBM is typically a sparse linear model (e.g., logistic regression or a single dense layer) that maps concept scores to the final prediction. This architectural choice guarantees that the contribution of each concept to the output is explicitly additive and independently inspectable. The learned weights directly quantify how much each concept influences the diagnosis—for example, a weight of +0.7 for 'spiculated mass margin' on a malignancy prediction. This contrasts sharply with deep networks where feature interactions are opaque and non-linear, making CBMs naturally compliant with regulatory explainability requirements under FDA SaMD guidelines.

A critical design challenge is ensuring the concept set is sufficiently complete to capture all predictive information. If important features are missing from the concept vocabulary, the model may either underperform or learn to encode residual information through concept leakage—where concept scores inadvertently encode unrelated features. For example, a 'bone lesion' concept might unintentionally correlate with patient age if not carefully regularized. Mitigation techniques include:

• Residual modeling that allows a small uninterpretable pathway alongside concepts
• Adversarial concept debiasing to remove spurious correlations
• Iterative concept set refinement with domain experts

Modern CBM variants relax the strict concept bottleneck to balance interpretability with performance. Concept Embedding Models learn dense representations of concepts rather than scalar scores, capturing richer semantics. Post-hoc CBMs first train a high-performance black-box model, then distill its knowledge into a concept-based explainable model. Self-explaining neural networks jointly optimize for prediction accuracy and concept prediction fidelity. These hybrid approaches address the accuracy-interpretability trade-off that pure CBMs sometimes face, making them viable for high-stakes medical imaging where both diagnostic precision and transparent reasoning are non-negotiable.

A method for interpreting neural network internal states by quantifying the model's sensitivity to user-defined high-level concepts. TCAV uses directional derivatives in the activation space to measure how strongly a concept like 'spiculated mass' influences a prediction, providing a quantitative bridge between latent space and human-understandable ideas without requiring architectural changes.

The approach of applying an interpretation method to a trained model after training is complete, without modifying the original architecture. Unlike Concept Bottleneck Models, which enforce interpretability by design, post-hoc methods like Grad-CAM or SHAP approximate a black-box model's reasoning. This distinction is critical for regulatory strategy: post-hoc methods audit existing models, while CBMs build auditability into the model itself.

A quantitative metric that evaluates whether an explanation accurately reflects the model's true reasoning process. It measures the correlation between attributed importance scores and the actual change in model output when features are perturbed. For CBMs, faithfulness is inherently higher because the reasoning chain—input → concepts → prediction—is the actual computational path, not an approximation.

The specific transparency and interpretability requirements mandated by health authorities like the FDA or under regulations such as the EU MDR. These standards demand that clinical AI decisions be auditable and validated for safety. Concept Bottleneck Models align naturally with regulatory expectations because their concept-based reasoning produces an auditable, human-readable decision trail that maps directly to clinical domain knowledge.

A human-AI collaboration paradigm where a medical professional actively reviews and interprets AI-generated explanations before making a final diagnosis. CBMs are particularly suited for this workflow because clinicians can intervene on mispredicted concepts—correcting an incorrectly identified 'irregular border' concept, for example—and observe how the final prediction updates, creating a true collaborative diagnostic process.

The false sense of security that arises from viewing a plausible-looking but ultimately unfaithful explanation. A saliency map may highlight anatomically reasonable regions while misrepresenting the model's actual decision logic. CBMs mitigate this risk because the explanation is the computation: the concept scores are the sole input to the final classifier, eliminating the gap between explanation and mechanism.