Generating Realistic Brain MRIs via a Conditional Diffusion Probabilistic Model (2212.08034v2)

Abstract: As acquiring MRIs is expensive, neuroscience studies struggle to attain a sufficient number of them for properly training deep learning models. This challenge could be reduced by MRI synthesis, for which Generative Adversarial Networks (GANs) are popular. GANs, however, are commonly unstable and struggle with creating diverse and high-quality data. A more stable alternative is Diffusion Probabilistic Models (DPMs) with a fine-grained training strategy. To overcome their need for extensive computational resources, we propose a conditional DPM (cDPM) with a memory-efficient process that generates realistic-looking brain MRIs. To this end, we train a 2D cDPM to generate an MRI subvolume conditioned on another subset of slices from the same MRI. By generating slices using arbitrary combinations between condition and target slices, the model only requires limited computational resources to learn interdependencies between slices even if they are spatially far apart. After having learned these dependencies via an attention network, a new anatomy-consistent 3D brain MRI is generated by repeatedly applying the cDPM. Our experiments demonstrate that our method can generate high-quality 3D MRIs that share a similar distribution to real MRIs while still diversifying the training set. The code is available at https://github.com/xiaoiker/mask3DMRI_diffusion and also will be released as part of MONAI, at https://github.com/Project-MONAI/GenerativeModels.

• The paper introduces a conditional diffusion probabilistic model that synthesizes high-quality 3D brain MRIs with fewer resources than traditional GANs.
• It employs a 2D conditional framework using a U-Net with self-attention to capture inter-slice dependencies and produce anatomically consistent images.
• Experimental results demonstrate superior performance across metrics such as FID and MS-SSIM, validating its potential for advanced medical image synthesis.

Conditional Diffusion Probabilistic Model for 3D MRI Synthesis

The paper "Generating Realistic Brain MRIs via a Conditional Diffusion Probabilistic Model" addresses the significant challenge of synthesizing high-quality 3D brain MRIs with limited computational resources and training data. The authors propose a Conditional Diffusion Probabilistic Model (cDPM) as an alternative to the commonly used Generative Adversarial Networks (GANs), which are known for instability and limited data diversity.

Introduction and Motivation

Medical image synthesis, particularly of brain MRIs, poses inherent challenges due to the high-dimensional nature of the data and limited availability of training samples. While GAN-based models have been explored, their instability and resource-intensive training process pose considerable limitations. This paper introduces cDPMs as a viable alternative, leveraging the more stable diffusion probabilistic models (DPMs) but integrating a conditional mechanism to efficiently generate 3D MRI volumes.

Methodology

The paper outlines the implementation of a 2D cDPM that synthesizes MRI subvolumes conditioned on subsets of slices from the same MRI. This approach allows the model to efficiently learn slice interdependencies, even when the slices are spatially disparate. Training involves learning these dependencies with an attention network, enabling the generation of anatomically consistent 3D brain MRIs.

Diffusion Probabilistic Model Framework

The authors first describe the underlying diffusion probabilistic model, which consists of two primary processes:

• Forward Diffusion Process (FDP): Transforms the real data distribution into Gaussian noise through a Markov chain.
• Reverse Diffusion Process (RDP): Reconstructs realistic data from noise, traversing backwards through the Markov chain.

With cDPMs, the conditioning of generation on existing slices allows the model to map noise to high-fidelity images without needing extensive computational resources.

Neural Architecture and Training

The cDPM utilizes a U-Net architecture enhanced with self-attention mechanisms to capture dependencies across slices. By iteratively generating new slices conditioned on previously generated ones, the model constructs a complete 3D MRI through multiple generation stages.

Experimental Results

Experiments conducted on t1-weighted brain MRIs from multiple datasets demonstrate the efficacy of cDPMs. The conditional architecture shows significant improvements over existing methods in synthesizing high-quality and anatomically diverse MRI scans.

• Qualitative Analysis: The synthesized MRIs are sharper with more defined gray matter boundaries compared to those generated by six other generative frameworks.
• Quantitative Measures: cDPM exhibits superior performance in metrics such as Maximum-Mean Discrepancy (MMD) and Fréchet Inception Distance (FID) across axial, coronal, and sagittal views.

The evaluation also includes MS-SSIM to assess diversity, where cDPM achieves scores closely aligning with real data, indicating its capacity to generate a wide range of realistic samples.

Implications and Future Directions

The research contributes a significant advancement in the field of medical image synthesis by providing a method that can generate high-quality 3D MRIs with reduced resource dependency. Practically, cDPMs can aid in training deep learning models where datasets are limited, enhancing their robustness and generalization.

Theoretically, the framework opens avenues for exploring diffusion-based models for other high-dimensional medical imaging modalities. Further, combining cDPM with advanced attention mechanisms might improve the model's ability to capture intricate anatomical structures, thus broadening the scope of its applications in medical diagnostics and treatment planning.

In conclusion, this work underscores the potential of integrating conditional mechanisms within diffusion probabilistic frameworks, offering an efficient, stable, and effective alternative for synthesizing complex medical imaging data.