CT Foundation Models Distilled into Editable Concept Bottlenecks for Lung Nodule Malignancy Prediction

A new study on arXiv (2608.07857) presents a method that improves how we understand predictions made by CT foundation models by transforming them into concept bottleneck models. This approach connects two fixed representations from the CT foundation model to eight characteristics of pulmonary nodules defined by radiologists, allowing for malignancy predictions based on these attributes and the size of the nodules. The research utilized CT-FM, a self-supervised encoder for 96^3-voxel nodules, and FMCIB, a contrastive encoder for 50-mm crops, developing eight ridge-regression concept heads using 2,610 LIDC-IDRI nodules. Models were trained on LUNA25 and tested against an internal set and the external DLCS cohort, with performance evaluated through R^2 and AUROC metrics. This work aims to improve AI interpretability in medical imaging for lung nodule malignancy predictions.