XCAT-GAN

XCAT-GAN for Synthesizing 3D Consistent Labeled Cardiac MR Images on Anatomically Variable XCAT Phantoms

Sina Amirrajab, Samaneh Abbasi-Sureshjani, Yasmina Al Khalil, Cristian Lorenz, Juergen Weese, Josien Pluim, Marcel Breeuwer

Accepted for 23rd International Conference on Medical Image Computing & Computer Assisted Intervention (Oral presentation) MICCAI 2020

Paper arXiv

Aim

To address the lack of properly annotated data by synthesizing anatomically meaningful, controllable, and variable CMR images suitable for medical data augmentation.

Propose

To create a virtual population of anatomically variable patients by leveraging XCAT phantoms and synthesize CMR images by learning modality-specific features through conditional GANs.

Framework

1. (simulated XCAT, XCAT labels) pairs to train a multi-class U-net.
2. (real images, segmentation) pairs to train a conditional GAN.
3. (synthetic XCAT, XCAT labels) to augment and replace the real images.

Results

• 3D volumetric consistency is achieved by using more XCAT labels.
• Data augmentation with XCAT-GAN synthetic images improves the generalizability of the segmentation network.
• Real data reduction experiment suggests to reduce the amount of real data up to 20% while retaining the performance.

Framework

Example Results

• 4-class vs 8-class XCAT-GAN synthesis

• Short axis view, end-diastolic phase and end-systolic phase