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

TL;DR - A prior-guided concept bottleneck model reduces the concept-annotation burden for interpretable cancer imaging while retaining diagnostic performance near black-box models.

• Combines limited annotations, distribution matching on unlabeled patients, and prior initialization of the diagnosis head.
• At 10% annotation, concept AUC rose from 0.619 to 0.741 for masses, 0.650 to 0.787 for calcifications, and 0.597 to 0.642 for pulmonary nodules.
• Ablations identify prior initialization as the primary contributor to improved concept detection.
• Zero-shot vision-language models remained unreliable for fine-grained tumor-level concept prediction.