Improving Heart Transplant Rejection Classification Training using Progressive Generative Adversarial Networks

Cardiac allograft rejection is a life-threatening complication that can occur in patients following heart transplantation. Endomyocardial biopsies, the current gold-standard for monitoring rejection, require manual identification of samples by experts; however, this can be subjective, costly and time-consuming. Computer-aided diagnosis models can potentially provide an automated analysis that offers accurate and consistent detection of biopsy samples. Unfortunately, the lack of a large dataset of rejection pathology signs, due to time consuming clinical annotations, limits the classification performance of such conventional AI-based models. In this paper, we developed a generative adversarial network (GAN) that creates synthetic tissue tiles from heart transplant whole slide images (WSIs) to serve as data for training rejection classifier models. To generate synthetic rejection histology regions, we used inspirational image generation (IIG) with a single rejection reference image. Additionally, to demonstrate objective improvement in image classification using synthetic rejection regions, we use a pretrained VGG-19 classifier to differentiate between rejection versus nonrejection tiles. We greatly improved classification performance, achieving an increase in the Matthews correlation coefficient (MCC) from 0.411 to 0.790 when the training set was augmented with our synthetic rejection tiles. Our model was able to create visibly realistic rejection tissue tiles which was used to augment the rejection tile database and enhanced automated rejection detection.