<p>Infrared thermography is a non-invasive imaging technology that detects spatial thermal patterns. It is widely used for identifying diseases and other health abnormalities as many of them result in variation of the thermal body response, which makes the technology attractive due to speed, low cost, and simplicity of the imaging procedure. However, analysis of such images is not trivial and relies on either human expertise or semiautomatic algorithms based on thermal gradients. Deep learning can solve this problem efficiently, however most of the deep learning-based methods currently developed for thermographic image analysis are based on discrimination models and hence require a large amount of labeled data for both healthy and unhealthy classes. The issue can be overcome by using anomaly detection methods based on Variational Autoencoder (VAEs) architecture. VAESIMCA is an anomaly detection method, which combines the capabilities of VAE and statistical methods for defining the anomaly detection rules based on the target class only utilized in Soft Independent Modelling of Class Analogy (SIMCA). In this paper the method was applied for classification of thermographic images based on two open-access datasets resulting in a strong performance with 95–100% efficiency. The paper covers both theoretical and practical aspects of using VAESIMCA for thermographic images as well as reports on the classification results in detail.</p>