Uncertainty-based feature selection for improved adequacy of dermoscopic image classification

In computer vision systems, the final measurement result can be a digital output provided by the software, i.e., a diagnostic value along with an uncertainty interval. In line with previous work on melanoma disease, we present here a platform for image analysis based on variational autoencoders (VAEs), a category of deep-learning generative models that learns how to reproduce an image in output by an encoder/decoder strategy. Latent variables (LVs) extracted from the VAE architecture are compact representations of the appearance of image objects and are used as phenotypes of malignant skin lesions. In this work, we go ahead with the state-of-the-art approaches by proposing a way to use the propagated uncertainty through the VAE system, related to skin tone variation and gel bubble effects, as an index to select more robust descriptors of melanoma malignancy, under the assumption that the less is the expected variation of a descriptor to perturbations, the more is its capability to discriminate it. Here, we consider images from the public dataset of dermoscopic images ISIC 2016-2017, 500 from malignant and 500 from benign findings, to address the binary classification problem of malignancy assessment. We reported the results of the uncertainty-based feature selection strategy using seven different standard classification methods, obtaining average accuracy (ACC) values of 0.82 (0.01) versus the case of 0.74 (0.05) obtained without any feature selection approach, demonstrating the increased adequacy of the model in diagnosing melanoma. General validity has been shown by comparing the results with those obtained using a standard transfer learning strategy based on convolution neural network (CNN). The proposed approach not only demonstrates an increased ACC in identifying malignant melanomas but also presents a completely novel approach for feature selection, framed within the theory of uncertainty propagation, specifically relevant in the case of black-box descriptors such as those extracted through deep learning.