EVOLUTION OF THE EFFECTIVENESS OF СONDITIONAL GAN FOR THE SYNTHESIS OF CT IMAGES OF THE BRAIN IN THE DIAGNOSIS OF STROKE

Abstract

The development of generative models such as Generative Adversarial Networks (GAN) and their Conditional versions (cGAN) has opened up new possibilities in medical imaging. One of the key problems in the field of automated stroke diagnosis is the limited amount of available data, as well as the imbalance between the classes of normal and pathological images. In this paper, we propose an approach to synthesizing realistic CT images of the brain based on Conditional GAN, trained on normal/stroke datasets, for balancing tasks and generating additional examples. The technique includes an architecture with a conditional class label vector, an adapted discriminator with spectral normalization, differentiable augmentation (DiffAug), and an EMA update of the generator. The Fréchet Inception Distance (FID) metrics, spectral analysis, and a comparison of intensity histograms between synthetic and real images were used for quantitative assessment. The obtained FID value of 194.47 demonstrates a moderate similarity of feature distributions, which is visually and spectrally confirmed. The results show that the conditional GAN is able to reproduce the anatomical and textural features of CT images of the brain, preserving brightness distributions and spatial patterns. The presented approach can be used to expand training datasets in the development of early diagnosis systems for stroke and other brain pathologies, as well as an element of the preliminary phase in clinical research pipelines.