When we look at a Magnetic Resonance Imaging (MRI) scan, we typically see a grayscale image mapping the structural anatomy of a patient. This visual representation is a magnitude image. However, the underlying physics of nuclear magnetic resonance does not produce real-valued scalars. Instead, raw data is acquired in the frequency domain—known as k-space—as a collection of complex numbers containing both magnitude and phase.
Traditional deep learning pipelines in medical imaging routinely discard the phase component or treat the real and imaginary channels as independent, real-valued entities. This practice introduces a fundamental information bottleneck. By shifting toward native complex-valued processing, recent research demonstrates how preserving the mathematical integrity of phase information unlocks superior image reconstruction, eliminates artifacts, and uncovers hidden physiological insights.
The Information Hidden in Phase
Why does phase matter if clinicians ultimately look at magnitude images? The phase component of a complex-valued signal is not noise; it is a direct reflection of physical and biological variations within the scanner.
In structural and functional MRI (fMRI), phase variations capture local magnetic field inhomogeneities, blood oxygenation levels (the BOLD effect), and the precise orientation of small blood vessels. As Kuang et al. observe in their work on multi-subject fMRI data, discarding the phase component throws away significant biological insights regarding tissue properties.
Raw Signal (Complex) ───► Magnitude (Structural Anatomy)
└───► Phase (Fluid Flow, Oxygenation, Field Variations)
Phase also tracks the timing and synchronization of neural signals across different brain regions. By building a framework that explicitly accounts for both magnitude and phase, Li et al. developed a Complex-valued Transfer Entropy (CTE) method that maps directed brain connectivity with drastically higher accuracy than magnitude-only approaches. Similarly, in multi-subject tensor decompositions, Kuang et al. demonstrated in pcsCPD that leveraging the unique "small phase" signature of task-related brain activity makes it possible to separate true neural activation from non-biological noise.
The Failure of Real-Valued Channels
A common workaround in deep learning is to feed complex data into standard convolutional neural networks (CNNs) by splitting the real ($R$) and imaginary ($I$) components into two independent input channels. While this allows standard architectures to run, it fails fundamentally to preserve the phase relationship.
A standard real-valued convolution mixes channels using independent weights, ignoring the algebraic rules of complex multiplication:
$$(R + iI)(W_R + iW_I) = (RW_R - IW_I) + i(RW_I + IW_R)$$
When a network treats these components as unrelated channels, it loses the geometric constraints of the complex plane, such as phase coherence and rotation. As Virtue et al. argue in Better than real, this independent treatment leads to suboptimal tissue parameter estimation in quantitative applications like MRI fingerprinting.
Furthermore, when accelerating scans by undersampling k-space, magnitude-only reconstructions become highly unstable. As Huang et al. show in a comparative study on cardiac diffusion tensor imaging, relying strictly on magnitude data causes models to introduce dangerous "hallucinations" at high acceleration rates (e.g., 8x undersampling), whereas preserving the full complex signal stabilizes the reconstruction.
Designing Native Complex Architectures
To prevent this information loss, researchers have designed networks that operate entirely within the complex domain, enforcing complex arithmetic throughout the forward pass.
Complex Input ──► Complex Convolution ──► Complex Activation (e.g., Cardioid) ──► Phase-Coherent Output
This architectural shift requires reimagining standard deep learning components:
- Complex Activations: Standard ReLUs can zero out negative real values while ignoring the imaginary component, destroying phase. To solve this, Virtue et al. proposed a "cardioid" activation function that scales magnitude based on phase rather than performing a hard thresholding on the real axis.
- Complex Convolutions: By binding the real and imaginary weights through complex matrix multiplication, the network maintains phase coherence across layers. Feng et al. expanded this with Dual-Octave Convolutions, which process real and imaginary components simultaneously across multiple spatial frequencies, outperforming real-valued models in parallel MRI reconstruction.
- Spatio-Temporal Processing: In dynamic settings like cardiac imaging, phase tracks motion over time. Küstner et al. developed CINENet, which utilizes complex-valued 4D spatio-temporal convolutions to reconstruct high-resolution cardiac motion from highly accelerated scans, preserving the physical signal integrity across both space and time.
Bypassing Reconstructions via k-Space
If the ultimate goal of an AI model is diagnostic analysis (such as classification or segmentation) rather than human viewing, converting raw complex frequency data into a magnitude image is an unnecessary and destructive step.
Raw k-Space (Complex) ──[Traditional Pipeline]──► Magnitude Image ──► Downstream AI (Phase Lost)
Raw k-Space (Complex) ──[Direct Pipeline]───────► KMAE Transformer ──► Downstream AI (Phase Preserved)
To exploit this, Li et al. introduced KMAE, a Transformer-based architecture that performs cardiac disease classification and tissue segmentation directly from raw k-space data. By bypassing the image reconstruction step entirely, the model preserves the raw phase information. The authors demonstrated that this direct frequency-domain processing makes the network remarkably robust against severe data loss, matching the performance of image-domain methods even when the scanner collects only 12.5% of the required data.
Overcoming the Magnitude-Only Data Legacy
Despite the clear advantages of complex-valued processing, a practical challenge remains: decades of historical clinical data have been saved exclusively as magnitude-only images, discarding the raw phase. This creates a data scarcity problem when trying to train modern generative priors that require complex-valued targets.
To bridge this gap, Luo et al. developed a workflow for training generative priors using phase augmentation. By taking readily available magnitude-only datasets and mathematically augmenting them with synthetic phase profiles, the authors successfully trained complex-valued generative models.
Their findings revealed a striking property of neural network optimization: generative priors trained on these phase-augmented complex images significantly outperform priors trained strictly on magnitude data. Forcing the network to learn the relationship between magnitude and a phase constraint—even a synthetic one—acts as a powerful regularizer, leading to higher reconstruction fidelity and fewer artifacts when deployed on real, undersampled complex scans.
Key Foundational Work
Adaptive independent vector analysis for multi-subject complex-valued fMRI data (2017) — contributed 1 concept elements to this direction.
Compressed Sensing MRI Using a Recursive Dilated Network (2018) — contributed 1 concept elements to this direction.
Better than real: Complex-valued neural nets for MRI fingerprinting (2017) — contributed 1 concept elements to this direction.