PhaseGen: A Diffusion-Based Approach for Complex-Valued MRI Data Generation

TL;DR

This paper addresses the limited use of complex-valued MRI raw data by introducing PhaseGen, a diffusion-based model that generates synthetic complex k-Space data conditioned on magnitude images. By preserving magnitude while synthesizing plausible phase, PhaseGen enables pretraining of models that require k-Space information and improves downstream tasks such as skullstripping in the frequency domain and MRI reconstruction with limited real data. The key findings show that training with synthetic phase data significantly enhances skullstripping generalization (DSC improving from 41.1% to 80.1%) and can boost reconstruction performance when combined with modest amounts of real data, approaching the performance of models trained on full real-phase data at around 15–20% real data. This work demonstrates a practical pathway to bridge magnitude-based datasets and complex-valued MRI data, potentially reducing data collection burdens while expanding the applicability of CVNNs to MRI tasks, with publicly available code for broader adoption.

Abstract

Magnetic resonance imaging (MRI) raw data, or k-Space data, is complex-valued, containing both magnitude and phase information. However, clinical and existing Artificial Intelligence (AI)-based methods focus only on magnitude images, discarding the phase data despite its potential for downstream tasks, such as tumor segmentation and classification. In this work, we introduce $\textit{PhaseGen}$, a novel complex-valued diffusion model for generating synthetic MRI raw data conditioned on magnitude images, commonly used in clinical practice. This enables the creation of artificial complex-valued raw data, allowing pretraining for models that require k-Space information. We evaluate PhaseGen on two tasks: skull-stripping directly in k-Space and MRI reconstruction using the publicly available FastMRI dataset. Our results show that training with synthetic phase data significantly improves generalization for skull-stripping on real-world data, with an increased segmentation accuracy from $41.1\%$ to $80.1\%$, and enhances MRI reconstruction when combined with limited real-world data. This work presents a step forward in utilizing generative AI to bridge the gap between magnitude-based datasets and the complex-valued nature of MRI raw data. This approach allows researchers to leverage the vast amount of avaliable image domain data in combination with the information-rich k-Space data for more accurate and efficient diagnostic tasks. We make our code publicly $\href{https://github.com/TIO-IKIM/PhaseGen}{\text{available here}}$.

Figures (7)

• Figure 1: Overview of the proposed method. Publicly available magnitude image domain data is used to generate synthetic complex-valued k-Space data. This synthetic data facilitates pretraining of models for clinical downstream tasks, which can later be fine-tuned using real-world data.
• Figure 1: Training
• Figure 2: Graphical representation of the complex-valued forward and reverse diffusion process. The input $z_t$ consists of magnitude and phase in the image domain. The added complex-valued noise primarily affects the phase while preserving magnitude.
• Figure 3: Example outputs of PhaseGen compared to original data. From left to right: input magnitude image, corresponding original phase, PhaseGen-predicted phase, unwrapped original phase, unwrapped predicted phase.
• Figure 4: Comparison of different phase data generation methods on the task of skullstripping directly in the frequency domain. The model trained with the data generated by the proposed methods outperforms the other methods in both metrics, DSC and HD, showing superior segmentation performance.