OrganLens: Self-Supervised Organ-Specific CT Representations

OrganLens is a novel self-supervised method for learning organ-specific representations from CT volumes. Unlike existing CT foundation models that produce a single volume-level representation, OrganLens conditions a shared encoder on a selected organ identity, then uses organ-specific distillation and anatomy-mask supervision to shape features for anatomy-weighted pooling. This approach preserves clinically relevant surrounding context, which is lost when encoding pre-separated organ volumes. At inference, the shared model produces distinct representations per organ, enabling targeted analysis for abnormalities, prognosis, or longitudinal change. The method addresses limitations of anatomy-aware approaches that either remove context or fail to condition the encoder before feature formation. OrganLens is introduced in a paper on arXiv (2607.25164) as a cross submission.