An Adaptive, Disentangled Representation for Multidimensional MRI Reconstruction

TL;DR

This work tackles multidimensional MRI reconstruction under data scarcity by learning a disentangled latent representation that separately encodes geometry and contrast via $oldsymbol{z}_g$ and $oldsymbol{z}_c$, with a style-based decoder and FiLM conditioning. A latent diffusion prior constrains these latents, and a zero-shot, self-supervised refinement adapts pretrained representations to target data, enabling robust reconstruction for accelerated $ ext{T}_1$ and $ ext{T}_2$ parameter mapping. Reconstruction combines data consistency with priors in a gradient-based optimization over $oldsymbol{z}_g$, $oldsymbol{z}_c$, and a refinement network $oldsymbol{ heta}_N$, using DDIM steps and a StochasticResample strategy to maintain stability. Experiments demonstrate superior accuracy and image quality compared to joint sparsity and zero-shot SSDU baselines, while illustrating the effectiveness of geometry–contrast disentanglement and the potential for subspace integrations in $ ext{T}_2$ mapping.

Abstract

We present a new approach for representing and reconstructing multidimensional magnetic resonance imaging (MRI) data. Our method builds on a novel, learned feature-based image representation that disentangles different types of features, such as geometry and contrast, into distinct low-dimensional latent spaces, enabling better exploitation of feature correlations in multidimensional images and incorporation of pre-learned priors specific to different feature types for reconstruction. More specifically, the disentanglement was achieved via an encoderdecoder network and image transfer training using large public data, enhanced by a style-based decoder design. A latent diffusion model was introduced to impose stronger constraints on distinct feature spaces. New reconstruction formulations and algorithms were developed to integrate the learned representation with a zero-shot selfsupervised learning adaptation and subspace modeling. The proposed method has been evaluated on accelerated T1 and T2 parameter mapping, achieving improved performance over state-of-the-art reconstruction methods, without task-specific supervised training or fine-tuning. This work offers a new strategy for learning-based multidimensional image reconstruction where only limited data are available for problem-specific or task-specific training.

Figures (8)

• Figure 1: A disentangled representation (using “geometry” and “contrast” features as examples): Once learned, geometry and contrast latents can be sampled from respective distributions and combined to generate images with target geometry or contrast.
• Figure 2: (a) The proposed representation learning strategy. Image transfer losses are used to train the disentangled representation. The decoder combines geometry latents from $\mathbf{X}_{g_i,c_i}$ with contrast latents from $\mathbf{X}_{g_j,c_j}$ to generate a new image (see right), which can be interpreted as contrast transfer for $\mathbf{X}_{g_i,c_i}$ or geometry transfer for $\mathbf{X}_{g_j,c_j}$. (b) The FiLM block for feature combination. At resolution $l$, the encoder feature maps $\mathbf{F}_{enc}^l$ are transformed into $\mathbf{z}_c^l$ through global average pooling (GAP) and a fully connected (FC) layer. The resulting $\mathbf{z}_c^l$ is then split into modulation parameters $\boldsymbol{\gamma}_c^{l}$ and $\boldsymbol{\beta}c^{l}$, which are applied to modulate the decoder feature maps $\mathbf{F}_{dec}^l$ at the same level.
• Figure 3: Feature disentanglement achieved: High-quality $\text{T}_1$w (first) and $\text{T}_2$w (last) images can be generated by sampling a geometry latent $\mathbf{z}_{g_{1}}$ from our learned latent diffusion model and combining it with two different contrast latents $\mathbf{z}_{c_{T_1}}$ and $\mathbf{z}_{c_{T_2}}$, using the trained decoder. Interpolation between $\mathbf{z}_{c_{T_1}}$ and $\mathbf{z}_{c_{T_2}}$ produced images with consistent geometry but varying contrasts (middle images). Combining a different geometry latent with $\mathbf{z}_{c_{T_1}}$ produced the same $\text{T}_1$ contrast but with different anatomical features (bottom).
• Figure 4: Reconstructed images (top left panel) from experimental $\text{T}_1$ mapping data and the corresponding error maps (top right panel) at AF=6. Images acquired with different flip angles are shown in different columns while results from different methods in respective rows. Zoomed-in regions from the images are shown in the bottom panel with artifacts that were reduced by our method indicated by blue arrows. Mean-squared errors and SSIM values are shown in the the error maps. The proposed method achieved the lowest error and highest SSIM, with most noise-like residuals.
• Figure 5: $\text{T}_1$ mapping results at $6\times$ acceleration (AF=6; brain masked). The proposed method preserved spatial details better with the highest mapping accuracy, as shown in the estimated $\text{T}_1$ maps (middle), zoomed-in regions (top), and the corresponding error maps (bottom). The overall errors and SSIM values for $\text{T}_1$ maps are shown.