Deep kernel representations of latent space features for low-dose PET-MR imaging robust to variable dose reduction (2409.06198v1)

Abstract: Low-dose positron emission tomography (PET) image reconstruction methods have potential to significantly improve PET as an imaging modality. Deep learning provides a promising means of incorporating prior information into the image reconstruction problem to produce quantitatively accurate images from compromised signal. Deep learning-based methods for low-dose PET are generally poorly conditioned and perform unreliably on images with features not present in the training distribution. We present a method which explicitly models deep latent space features using a robust kernel representation, providing robust performance on previously unseen dose reduction factors. Additional constraints on the information content of deep latent features allow for tuning in-distribution accuracy and generalisability. Tests with out-of-distribution dose reduction factors ranging from $\times 10$ to $\times 1000$ and with both paired and unpaired MR, demonstrate significantly improved performance relative to conventional deep-learning methods trained using the same data. Code:https://github.com/cameronPain

• The paper introduces a novel deep kernel framework that combines MR-derived and PET latent features for robust low-dose PET image reconstruction.
• It employs mutual information constraints to enhance anatomical fidelity and reduce noise, validated by improved PSNR and SSIM at variable dose levels.
• The method processes both paired and unpaired PET-MR datasets, demonstrating flexibility for clinical applications in diverse imaging scenarios.

Deep Kernel Representations of Latent Space Features for Low-dose PET-MR Imaging Robust to Variable Dose Reduction

This paper introduces an innovative approach for low-dose PET image reconstruction by leveraging deep kernel representations of latent space features, explicitly targeting robustness across variable dose reductions. The primary motivation is to address the intrinsic challenges in low-dose PET, where reduced radioactive tracer dose leads to poorly conditioned data and thus noisy, diagnostically suboptimal images. Traditional deep learning methods often fail when dealing with out-of-distribution dose reduction scenarios or under clinical conditions not reflected in their training data.

Methodology

The authors propose a novel method integrating deep learning with kernel representations, specifically designed to handle varying dose levels by incorporating MR-derived information. The architecture consists of two primary encoding branches: one for MR and one for PET. These branches produce latent space features used to construct kernel functions. The kernel functions derived from MR data establish a robust representation, which, when combined with PET encodings, form kernel features that enhance image reconstruction.

Key contributions of the paper include:

1. Latent-space Kernel Functions: Construction of kernel matrices from MR-derived feature maps to regularize PET latent space features.
2. Information Constraints: Application of mutual information constraints on deep PET features to ensure these encode information not already captured by shallow MR-derived kernels.
3. Applicable to Paired and Unpaired Imaging: Demonstration of the method's efficacy with both paired and unpaired PET-MR datasets.
4. Empirical Validation: Extensive testing across varying dose reduction factors, showcasing superior performance relative to conventional deep learning and kernel methods.

Results

The method is evaluated on a cohort of healthy subjects, both for $^{18}F$-FDG and $^{18}F$-FDOPA PET brain images. The reconstruction performance is assessed for significantly reduced dose levels, ranging from $\times 10$ to $\times 1000$ reduction factors.

Performance Metrics:

• Paired PET-MR: For paired $^{18}F$-FDG and $^{18}F$-FDOPA datasets, the proposed method demonstrated robust performance across all dose reduction levels. Quantitative metrics showcased improved Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) compared to conventional benchmarks.
• Unpaired PET-MR: For unpaired datasets, where a template MR image is used, the method showed comparable robustness, though performance naturally varied depending on the coregistration accuracy of template MR data to individual PET scans.

The inclusion of information constraints showed significant improvement in anatomical preservation and reduced bias across varying dose levels, particularly at extreme dose reduction scenarios.

Implications and Future Directions

The implications of this method are both practical and theoretical. Practically, the approach allows for flexibility in clinical settings by ensuring robust image reconstruction even when the administered dose varies significantly. This is crucial for sensitive patient groups like pediatrics or those requiring frequent imaging. Theoretically, the integration of kernel methods within deep learning frameworks sets a precedent for hybrid approaches in medical imaging, merging the strengths of both paradigms.

Future development could explore:

1. Extension to Other Modalities: Adapting the approach for other anatomical regions or PET tracers.
2. Enhanced Regularization Techniques: Further refining kernel functions or incorporating other priors that may enhance deep feature encoding.
3. Scalability: Implementing fully 3D versions of the method to take full advantage of volumetric data.

In conclusion, the deep kernel representation of latent space features offers a significant advancement in robust low-dose PET image reconstruction. This approach balances high-quality image synthesis with adaptability to varying clinical conditions, potentially reshaping low-dose PET imaging methodologies.