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

Han Bao; Jie Deng; Shihao Xing; Yishan Zhong; Wenqi Shi; Benoit Marteau; Bibhuti Das; Bahig Shehata; Shriprasad Deshpande; May Dongmei Wang

Abstract

Heart Transplant Rejection (HTR) is an exceedingly uncommon condition that necessitates early detection for successful treatment and to prevent lasting damage to the transplanted heart. Unfortunately, the current diagnostic process is time-consuming and lacks consensus among medical professionals. Introducing an automated diagnosis pipeline would greatly streamline the clinical workflow, serving as an additional clinical decision support tool that offers a second opinion to enhance agreement among clinicians. Traditionally, developing an automated image analysis tool of this nature requires a substantial amount of labeled data. However, due to the rarity and inherent case imbalance of HTR, this task becomes particularly challenging. Our dataset comprises 1,614 rejection tile images and ~190 times more non-rejection tile images. To address the scarcity of real-world examples, we present a novel approach featuring synthetic image generation using a diffusion model, where synthetic images of rejection were generated. We conducted a comparative analysis of classification using the dataset both with and without synthetic rejection tiles. The introduction of synthetic augmentation resulted in an improvement in sensitivity from 0.781 to 0.981 and in AUROC from 0.984 to over 0.998.

Keywords: histopathology; generative model; whole slide image; diffusion

Links
[Full text PDF][Bibtex]