Kvit Achieves MRI Classification Directly from K-Space Data with Novel Patching

Scientists are tackling a significant limitation in medical image analysis: the loss of crucial phase information when processing Magnetic Resonance Imaging (MRI) data. Moritz Rempe, Lukas T. Rotkopf, and Marco Schlimbach, from the Institute for AI in Medicine (IKIM) and the Technical University Dortmund, alongside Helmut Becker, Fabian Hörst, and Johannes Haubold et al, present a novel complex-valued Vision Transformer (kViT) that classifies MRI scans directly from raw k-space data , bypassing traditional reconstruction methods. This innovative approach not only preserves vital phase details but also overcomes the limitations of conventional neural networks by leveraging the global characteristics of k-space, achieving competitive classification accuracy with substantially reduced computational demands and up to 68% less VRAM usage. The kViT therefore promises a paradigm shift towards faster, more efficient, and resource-accessible AI-powered MRI analysis directly at the point of scan.

The research team achieved this by designing an architecture specifically tailored to the unique characteristics of k-Space, moving beyond standard neural networks that struggle with its global, non-local nature. The team observed a reduction in Video RAM (VRAM) consumption during training of up to 68x compared to standard methods. Furthermore, the proposed physics-informed radial patching strategy overcomes the limitations of Cartesian grid patching in the frequency domain, allowing for a more accurate representation of k-Space data. Researchers hypothesized that standard convolutional or patch-based operations are ill-suited for the global, non-local characteristics inherent in raw frequency-domain data, motivating the development of kViT. This encoder utilizes self-attention mechanisms to model long-range dependencies within k-Space, effectively capturing the non-local relationships between data points. Furthermore, the research revealed a significant reduction in VRAM consumption during training, achieving up to a 68× decrease compared to standard methods. The radial k-patching strategy proved effective in capturing the non-local properties of k-space data, overcoming limitations associated with traditional Cartesian grid patching. Experiments revealed that kViT maintains robust classification performance even with high acceleration factors, indicating its resilience to data undersampling commonly used to reduce scan times. Data shows the model’s ability to process complex-valued inputs effectively, unlocking the full potential of k-space data for AI applications.

Specifically, the team evaluated performance on fastMRI Prostate, Knee, and in-house Glioma datasets, consistently demonstrating the model’s robustness across diverse anatomical regions and imaging protocols. Measurements confirm that the proposed architecture achieves competitive classification accuracy while drastically reducing computational costs, offering a paradigm shift in MRI data processing. The kViT employs a radial k-patching strategy, respecting the energy distribution within the frequency domain, and operates directly on complex k-space data, unlike standard methods. Importantly, kViT exhibits enhanced robustness at high acceleration factors and significantly reduces VRAM consumption during training, up to 68% less than standard methods, paving the way for efficient, direct-from-scanner AI analysis. Future research directions include exploring hybrid architectures combining k-space and image-space information, and investigating the benefits of pretraining models on larger datasets, despite the challenges of data scarcity in MRI raw data.