ATTRI-SSC-VAE: Multi-Attribute Regularized Sparse Coding VAEs for Interpretable Medical Image Representation

Xin Gao, Lu Wang, Xu Chenhao, 马立欣, He Dantong, Ye Luo

Keywords: Structured Sparse Coding, Variational Autoencoders (VAEs), Attribute Regularization, Interpretable Representation Learning, Controllable Generation, Explainable Medical Imaging

Abstract: Explainable image representations are critical in medical imaging, where interpretability is essential for both clinical trust and decision-making. We introduce Attri-SSC-VAE, a novel framework that extends Structured Sparse Coding Variational Autoencoders (SSC-VAEs) with attribute regularization and multi-attribute mapping. Our approach leverages sparse coding to discretize image representations into a dictionary of latent components while preserving generative flexibility through a VAE encoder–decoder structure. To enhance interpretability, we impose attribute regularization on the coding coefficients, explicitly associating dictionary elements with meaningful clinical attributes. Furthermore, a multi-attribute mapping mechanism enables disentanglement across attributes, ensuring that variations in specific coding coefficients correspond to consistent and explainable changes in image features. This property allows for controlled image editing, where manipulating the coefficients associated with target attributes results in semantically aligned modifications in generated images. Experiments on medical imaging datasets demonstrate that Attri-SSC-VAE not only achieves competitive reconstruction and generation performance but also provides interpretable, attribute-aware representations that improve trustworthiness and practical utility in clinical applications.

Primary Area: interpretability and explainable AI

Submission Number: 10546