Rare Heart Transplant Rejection Classification Using Diffusion-Based Synthetic Image Augmentation

Heart Transplant Rejection (HTR) is a rare condition that requires early detection to prevent lasting damage to the transplanted heart. Unfortunately, the current HTR grading through biopsy image classification lacks consistency among pathologists. In addition, it is a time-consuming task. In this work, we have developed an automated diagnosis pipeline to streamline the heart transplant histopathology image quantification and classification, in order to provide objectivity for clinical decision support for pathologists. Traditionally, developing an automated image classification requires a substantial amount of labeled data. However, HTR is a rare condition and the dataset is usually unbalanced. For example, the dataset from DNA Based Transplant Rejection (DTRT) comprises 1,509 rejection tile images and 190 times more non-rejection tile images. To address the small data sample challenge in training the classifiers, we developed a novel strategy that used diffusion model to generate synthetic images of rejection. We conducted comprehensive HTR grade classification comparing results using dataset with synthetic rejection tiles versus the dataset without any synthetic rejection tiles. The introduction of synthetic augmentation resulted in an improvement from 0.781 to 0.981 for sensitivity, and an improvement from 0.984 to over 0.998 in AUROC. This study illustrated that synthetic data augmentation is a feasible strategy in developing AI solutions for rare diseases. In the future, we will expand in this direction to benefit more rare disease clinical decision support development.