SPD-SKINNET: SPECTRAL RIEMANNIAN LEARNING WITH COVARIANCE REPRESENTATIONS FOR SKIN LESION CLASSIFICATION

Abstract

Automatic skin lesion diagnosis is a fundamental challenge in medi-

cal image analysis because of requiring accurate and interpretable rep-

resentations of complex visual patterns. Most existing deep learning

approaches rely on Euclidean feature embeddings, which primarily cap-

ture first-order statistics and often fail to model the intrinsic geometric

and textural structure of skin lesions. In this paper, we propose SPD-

SkinNet, a spectral–Riemannian deep learning framework that repre-

sents skin lesion features as symmetric positive definite (SPD) matrices,

enabling the modeling of second-order statistics via feature covariance.

By performing spectral decomposition on the SPD manifold, we extract

eigenvalues that serve as compact and interpretable descriptors of lesion

texture. The initial experimental results on the synthetic highlight the

potential of integrating Riemannian geometry and spectral analysis into

deep learning for medical imaging.