Comparison of the Effectiveness of Semi-Supervised Learning Methods Based on Autoencoders for Skin Cancer Photo Classification Tasks

Abstract

In this paper, we reviewed and compared the effectiveness of semi-supervised learning methods based on autoencoders for the task of classifying photos of skin cancers. The paper proposes a system for evaluating semi-supervised learning methods and suggests modifications to the architectures described in the literature. The relevance of this work is to study the effectiveness of using various semi-supervised learning methods on applied classification tasks, The effective use of semi-supervised learning methods for working with medical data can potentially significantly increase the quality of artificial intelligence in medicine, since unlabeled medical data is easier to obtain and does not require the involvement of expensive specialists. As the experiments conducted in this study have shown, semi-supervised learning methods based on autoencoders significantly improve the performance metrics of the model for the task of classifying photos of skin cancers. In our study, we considered a conventional convolutional autoencoder, a variational autoencoder, and a ladder neural network, all adapted for semi-supervised learning tasks. The HAM10000 dataset was used as the data for the experiments. For each model, 3 experiments were conducted with different amounts of labeled data in the training set (5, 10, 30 % of labeled data). The studies showed that the variational autoencoder demonstrates exceptional results on a large amount of labeled data, but does not adapt well to a small amount. The convolutional autoencoder is effective only for a large amount of labeled data and is significantly inferior to more complex architectures in all experiments. The study has shown that the ladder neural network is very effective for a small amount of labeled data compared to the other architectures considered. Thus, taking into account the specifics of medical data classification tasks, we can conclude that the ladder neural network is the most effective for this type of task.