CoLa-Diff: Conditional Latent Diffusion Model for Multi-Modal MRI Synthesis

1 Introduction

• MRI acquisition can be time-consuming, expensive, or infeasible for some modalities, while available modalities provide complementary information for brain-tumour management.Trans-modal synthesis maps available modalities to a missing target modality.Source: §1 · pp. 1–2
• Prior multi-modal synthesis methods outperformed single-modality translation, but traditional approaches relied heavily on alignment and were poorly scalable.GAN-based methods included MM-GAN, DiamondGAN, and ProvoGAN.Source: §1 · p. 2
• GANs can struggle to model complex, variable multi-modal data, and may require fine-tuning to avoid gradient vanishing and mode collapse [2].The cited passage also notes that GAN performance can be limited when data variability is greater.Source: §1 · p. 2
• Earlier diffusion MRI synthesis focused on one-to-one translation, while many-to-one synthesis requires balancing multiple conditions and faces memory, anatomical-consistency, and sampling-speed challenges.Image-domain diffusion retains all Markov states in memory, and noise randomness can disrupt target anatomy.Source: §1 · p. 2
• CoLa-Diff is presented as the first diffusion-based many-to-one MRI synthesis model, operating in latent space and aiming to preserve anatomy with accelerated sampling.The authors state, “As far as we know,” to qualify the first-model claim.Source: §1 · p. 2
• The bespoke latent-space architecture uses similar cooperative filtering to address risks from excessive information compression and high-dimensional noise.The architecture is designed to facilitate diffusion operations in latent space.Source: §1 · p. 3
• Brain-region masks guide each diffusion step to preserve anatomical structure, while auto-weight adaptation balances conditions to leverage relevant multi-modal information.The contribution passages describe structural guidance and automatic condition-weight adaptation as separate design elements.Source: §1 · p. 3

2 Multi-conditioned Latent Diffusion Model

• CoLa-Diff encodes the target image into latent features, then uses a network conditioned on compressed MRI contrasts and masks to predict noise and recover the features for decoding.The forward process adds noise over T steps; reverse diffusion repeatedly recovers features before decoding the final representation into a synthesis image.Source: §2 · pp. 3–4

2.1 Latent Space Network

• Residual convolution blocks, wider receptive-field fusion, and similar cooperative filtering address latent information loss and noise in downsampled features.The filtering decomposes features with a wavelet transform, groups components for block matching or thresholding, then applies an inverse transform to reconstruct denoised features.Source: §2.1 · p. 4, Similar Cooperative filtering · p. 4

2.2 Structural Guidance

• FSL-FAST segments white matter, grey matter, cerebrospinal fluid, and tumour; their masks and density distributions condition reverse diffusion to guide anatomical structure.The paper motivates this guidance by noting that diffusion models can corrupt anatomical structure and that brain structures overlap in MRI density distributions.Source: §2.2 · p. 5

2.3 Auto-weight adaptation

• Auto-weight adaptation adjusts encoded condition channels using embedding weights and learnable activation parameters to balance the multiple conditions during training.The stated aim is to maximize relevant information while minimizing redundant information across conditions.Source: §2.3 · pp. 5–6

3.1 Comparisons with State-of-the-Art Methods

• Evaluation uses BRATS 2018 and IXI, comparing CoLa-Diff with MM-GAN, Hi-Net, ProvoGan, and LDM across seven cases in two datasets.BRATS includes 285 glioma patients; IXI contains 200 healthy-brain scans. CoLa-Diff uses 1,000 diffusion steps and knowledge-distillation-based fast sampling.Source: Datasets and Baselines · p. 6, Implementation Details · p. 6, Quantitative Results · p. 7
• Up to 6.01 dB on PSNR and 5.74% on SSIM separate CoLa-Diff from other models; against each task’s best other model, maxima are 0.81 dB on PSNR and 0.82% on SSIM.Table 1 reports test-set mean±std for BRATS and IXI.Source: Quantitative Results · p. 7, Table 1 · p. 7
• CoLa-Diff better maintains anatomy, including complex sulci and tumour boundaries, in T1ce BRATS and PD IXI examples, with fewer errors in the maps.Figure 2 shows detail enlargements in rows 1 and 3 and corresponding error maps in rows 2 and 4.Source: Figure 2 · p. 7, Qualitative results · p. 8

3.2 Ablation Study and Multi-modal Exploitation Capabilities

• In the BRATS T1+T1ce+FLAIR→T2 ablation, auto-weight adaptation raises PSNR by 1.9450 dB and SSIM by 4.0808%.The study removes components individually and reports that each contributes to performance improvement.Source: §3.2 · p. 8, Table 2 · p. 8
• With more input modalities for T2 synthesis, PSNR rises from 26.6355 dB to 28.3126 dB in BRATS and from 32.164 dB to 32.8721 dB in IXI; SSIM’s maximum uplift is 1.9603.The tests progressively increase the input modalities.Source: §3.2 · p. 8, Table 2 · p. 8

• CoLa-Diff combines a bespoke network backbone, similar cooperative filtering, structural guidance, and auto-weight adaptation for multimodal MRI synthesis.Source: §4 · p. 8
• Experiments support state-of-the-art performance in multimodal MRI synthesis tasks.Source: §4 · p. 8
• The authors suggest CoLa-Diff could help reduce MRI scanning burden and benefit patients and healthcare providers.Source: §4 · p. 8