Ensemble System for Skin Cancer Detection Based on the Analysis of Heterogeneous Dermatological Data Using Multimodal Neural Networks

Currently, skin cancer is the most common cancer pathology in the human body and one of the leading causes of death in the world. Therefore, it is relevant to develop high-precision intelligent systems for auxiliary diagnostics of skin cancer in the early stages. Ensemble models are one of the promising methods for increasing the accuracy of intelligent classification systems by reducing the dispersion and variability of forecasts of individual components of the overall system. Multimodal architectures can significantly increase the accuracy of neural network classification through methods for parallel analysis of heterogeneous dermatological data. The work proposes ensemble intelligent systems for analyzing heterogeneous dermatological data based on multimodal neural networks with various convolutional architectures. The accuracy of the weighted average ensemble based on multimodal systems using convolutional architectures AlexNet, Inception_v4, Densenet_161 and ResNeXt_50 for multi-class classification was 86.88%, and for binary estimation the accuracy was 94.10%. The accuracy of the weighted ensemble based on similar multimodal systems with weights corresponding to the accuracy of each base classifier was 86.82% for the original multiclass classification and 93.82% for the binary evaluation. The developed ensemble systems can be implemented as a high-precision auxiliary diagnostic tool to help make a medical decision.