Flip Distribution Alignment VAE for Multi-Phase MRI Synthesis

TL;DR

This paper tackles multi-phase CE MRI synthesis by explicitly decoupling shared and phase-specific latent features using a lightweight variational autoencoder. It introduces Flip Distribution Alignment (FDA) to enforce symmetry of latent distributions around a standard normal and employs a Y-shaped bidirectional training strategy to enable both self-reconstruction and cross-phase synthesis with simple mean-flipping. The approach achieves superior synthesis quality while significantly reducing model size and inference time compared with state-of-the-art deep autoencoders, demonstrated on a liver MRI dataset with six synthesis tasks. These characteristics imply practical impact for faster, safer MRI workflows with limited annotated data. The method combines a compact VAE backbone, structured latent space constraints, and bidirectional training to deliver interpretable and efficient multi-phase image synthesis.

Abstract

Separating shared and independent features is crucial for multi-phase contrast-enhanced (CE) MRI synthesis. However, existing methods use deep autoencoder generators with low parameter efficiency and lack interpretable training strategies. In this paper, we propose Flip Distribution Alignment Variational Autoencoder (FDA-VAE), a lightweight feature-decoupled VAE model for multi-phase CE MRI synthesis. Our method encodes input and target images into two latent distributions that are symmetric concerning a standard normal distribution, effectively separating shared and independent features. The Y-shaped bidirectional training strategy further enhances the interpretability of feature separation. Experimental results show that compared to existing deep autoencoder-based end-to-end synthesis methods, FDA-VAE significantly reduces model parameters and inference time while effectively improving synthesis quality. The source code is publicly available at https://github.com/QianMuXiao/FDA-VAE.

Figures (5)

• Figure 1: Overview of existing AE-based medical image synthesis strategies. (a) Basic one-to-one autoencoder. (b) Multi-phase many-to-one autoencoder. (c) Autoencoder with latent space comparative learning. (d) Variational Autoencoder (VAE). (e) Our proposed method: Flip Distribution Alignment Variational Autoencoder (FDA-VAE).
• Figure 2: Overview of the proposed Flip Distribution Alignment Variational Autoencoder (FDA-VAE). The model consists of a shared encoder, two independent decoders, and a flip distribution alignment (FDA) constraint layer. During training, a pair of different phase MRI images is input to obtain self-reconstructed and cross-phase transformed outputs. In the inference stage, only the target decoder is retained. The image is encoded to obtain the latent distribution, and the mean vector is flipped before decoding, generating the target-phase image from the flipped distribution.
• Figure 3: Convergence process of input and target distributions: (a) KL divergence only, (b) KL divergence + FDA.
• Figure 4: Visualization of synthesis result and errors heat maps.
• Figure 5: FDA feature decoupling in pixel level (a) & latent space level (b).