Sensitivity analysis of latent variables in Variational Autoencoders for Dermoscopic Image Analysis

Abstract

The advances in the deep learning field have paved the way to novel strategies to represent digital image data in the form of synthetic descriptors. Variational Auto-Encoders (VAE) architectures are generative powerful tools not only to reconstruct input images but also to extract meaningful information for the task of pattern classification. The first part of the VAE network, called encoder, aims to condense the image information into a reduced set of low-level descriptors, called latent variables. The second part, called decoder, aims to use the latent variable in a reverse process that reconstructs the original image in output. In this work, we exploited the VAE-based latent representation of colour normalized dermoscopic images for the discrimination of malignant and benign skin lesions. In particular, we investigated the sensitivity to the effect of skin colour variations over the final reconstruction error and on the discrimination capability of the VAE latent variables in terms of individual Area Under the roC curve (AUC). By exploiting and adapting state-of-the art skin colour variation models we obtained a performance worsening of about 10% either in the reconstruction error and in the discrimination capability of the latent variables. The achieved preliminary results demonstrate that, with suitable VAE adaptation, latent descriptors could be used in automatic skin lesions classification frameworks.