An innovative medical image synthesis based on dual GAN deep neural networks for improved segmentation quality

Artificial intelligence networks, precisely deep learning, have emerged as a truly life-impacting potential in healthcare - particularly in the area of medical diagnosis - and have achieved ground-breaking results in data enhancement, augmentation and reconstruction. In this paper, we propose a novel pipeline of an architecture of Generative Adversarial Networks (GAN) for Improved Segmentation Quality to attain better detection, diagnosis and treatment of diseases. The proposed method consists in creating an extended medical image distribution that will not only augment the data but also contribute to the achievement of good results in semantic segmentation tasks for better computer-aided medical diagnosis. In fact, our method is based on a pipeline with two stages representing the Segmentation 2-GAN (Seg2GAN) architecture, for the synthetic data distribution generation to enhance the quality of input images, followed by the segmentation phase using the new data. To validate this approach, we demonstrate an efficient medical data synthesis for diverse segmentation structures: the blood vessels of the retinal and the coronary, and the knee cartilage. Quantitative evaluations are presented as accuracy, sensitivity, precision and a dice score for three datasets. The segmentation results show that using the new synthetic data improves the same model trained on original real images; e.g., the gain of dice score is around of 8%, 21% and 50% for the blood vessels of the retinal and the coronary and for the knee cartilage, respectively. These obtained results compete with the state-of-art on multiple performance metrics.