cHeartFlow: Synthesizing cardiac MR images from sketches

Xinrui Zu, Qian Tao

Keywords: Medical image synthesis, contrastive learning, cardiac MRI

Abstract: Medical image synthesis is a highly promising approach to generate and augment medical data, which suffers from high acquisition costs and stringent privacy restrictions. However, current generation methods typically require detailed anatomical annotations and are limited in generating high-quality, anatomy-compliant images. To overcome the limitations, we present contrastive HeartFlow (cHeartFlow), a novel generative framework to synthesize cardiac magnetic resonance (CMR) images from simple sketches by training on contrastive pairs of images and sketches. cHeartFlow supports one-step synthesis and allows multi-step synthesis for a flexible trade-off between faithfulness and realism. We illustrate the effectiveness and generalizability of cHeartFlow through our experiments on different input sketches, compared with GAN-based and diffusion-based baselines.

Submission Number: 60