Less Annotation, More Interpretation: Prior-Guided Concept Bottleneck Models for Interpretable Cancer Imaging Diagnosis

Deep learning models in cancer imaging often achieve high diagnostic accuracy, but their "black-box" nature hinders clinical trust because they do not explain which radiological features drive their predictions. Concept bottleneck models (CBMs) solve this transparency problem by forcing networks to first predict human-understandable radiological concepts (such as lesion shape, margin characteristics, or spiculation) before using those concepts to diagnose malignancy. However, standard CBMs require extensive, expensive instance-level expert concept annotations for every training image, limiting their practical deployment in clinical settings.

To address this annotation bottleneck, the authors investigate how to train interpretable CBMs when concept labels are scarce. They evaluate partial supervision, zero-shot vision-language models, and a novel hybrid approach that leverages clinical and statistical priors to guide concept learning and stabilize training when ground-truth labels are largely missing.

The authors propose a prior-guided hybrid CBM that integrates three key components:

1. Partial Instance Supervision: For the small fraction of patients with expert concept annotations, standard binary cross-entropy loss is applied directly to the predicted concept probabilities.
2. Class-Conditional Prior Distribution Matching: For unannotated patients, a distribution matching loss compares the mean predicted concept vectors against expected class-conditional concept prevalence priors (derived either manually from guidelines or empirically from the training dataset).
3. Prior Initialization and Regularization: The concept-to-diagnosis linear head is initialized using logit-based probabilities derived from the clinical priors, mapping concepts to malignancy in a clinically plausible way from the start. An optional L2 regularization term anchors the weights to this initialization.

The authors evaluate this method on three benchmark datasets: CBIS-DDSM mammographic masses, CBIS-DDSM calcifications, and LIDC-IDRI pulmonary CT nodules, varying concept annotation fractions from 0% to 100%.