Alignment of Self‐Supervised Learning Representations With Radiomic Features in Multiphase Renal Computed Tomography

Self‐supervised learning (SSL) has emerged as a promising approach in medical image analysis, offering the ability to learn robust feature representations from unlabeled data. However, the extent to which these embeddings align with clinically interpretable features remains largely unaddressed. This study aims to assess the degree to which SSL embeddings encode radiomics‐aligned, interpretable features within multiphase renal computed tomography (CT).

We analyzed four‐phase contrast‐enhanced CT scans of renal tumors, including the noncontrast, corticomedullary, nephrographic, and excretory phases. Radiomic features included first‐order statistics, texture‐based features, and shape descriptors. SSL embeddings were obtained from three models: Simple framework for contrastive learning of visual representations (SimCLR), distillation with no labels (DINO)—a self‐distillation with no labels method, and bootstrap your own latent (BYOL). Feature‐level alignment was quantified using pairwise Spearman correlation analyses between SSL embeddings and radiomic features.

DINO demonstrated the highest alignment with radiomic features, particularly texture‐based features, such as gray‐level size zone matrix (GLSZM) (up to 22%) and neighboring gray tone difference matrix (NGTDM) (up to 35%), followed by SimCLR. BYOL showed minimal alignment across all feature families. Among imaging phases, the nephrographic phase exhibited the strongest correlations overall.

These findings indicate that SSL models, particularly DINO and to a lesser extent SimCLR, are capable of encoding patterns that align with established radiomic descriptors, especially texture‐based features. This partial alignment underscores the potential of incorporating radiomics‐informed metrics into the evaluation framework for SSL models, enabling more interpretable and clinically relevant model selection in medical imaging applications.