A MATHEMATICAL MODEL OF HYBRID CNN–GAN ARCHITECTURES FOR AUTOMATED DIAGNOSTICS UNDER DATA LIMITED CONDITIONS
TL;DR
The proposed approach generates additional synthetic images for each real positive example, thus compensating for class imbalance and improving classification accuracy under limited data conditions, and provides a foundation for optimizing hybrid architectures in few-shot learning scenarios.
Abstract
This paper addresses the problem of automated diagnosis of rare oncological diseases using medical imaging data with hybrid architectures combining Convolutional Neural Networks and Generative Adversarial Networks. The study focuses on a rigorous mathematical formalization of the entire pipeline, including data structure description, stratification procedures, preprocessing methods, and synthetic image generation to augment training datasets. The proposed approach generates additional synthetic images for each real positive example, thus compensating for class imbalance and improving classification accuracy under limited data conditions. The presented framework eliminates the ambiguity typical of purely verbal problem formulations, ensures reproducibility of results, and provides a foundation for optimizing hybrid architectures in few-shot learning scenarios. Future research directions include integrating the proposed approach into clinical workflows, investigating the impact of synthetic data parameters on diagnostic performance, and adapting the methodology to various types of rare pathologies.