Translating MRI to PET through Conditional Diffusion Models with Enhanced Pathology Awareness

Technical Deep Dive: Translating MRI to PET through Conditional Diffusion Models with Enhanced Pathology Awareness

Overview

The presented work, called PASTA (Pathology-Aware croSs-modal TrAnslation), is a novel framework for translating 3D brain MRI scans to corresponding PET images using conditional diffusion models. The key contributions are:

• A symmetric dual-arm architecture that strongly conditions the denoising process on the input MRI and clinical data, enabling high-quality PET synthesis with enhanced pathology awareness.
• Integration of multi-modal conditions through adaptive normalization layers to facilitate pathology-preserving PET generation.
• A cycle exchange consistency training strategy to further lift the translation quality.
• A memory-efficient volumetric generation strategy for producing consistent 3D PET scans.

The proposed PASTA framework significantly outperforms state-of-the-art GAN and diffusion-based translation methods, both quantitatively and qualitatively. In Alzheimer's disease diagnosis, PASTA's synthesized PET scans achieve an AUC that nearly matches actual PET, surpassing MRI by over 4%.

Problem & Context

• Positron emission tomography (PET) is a critical functional imaging technique for diagnosing neurodegenerative diseases like Alzheimer's, but its high costs and radiation exposure limit widespread use.
• In contrast, magnetic resonance imaging (MRI) is more accessible but less sensitive for such diagnostic purposes.
• To address this limitation, generating synthetic PET from MRI data is a promising approach, but faithfully preserving pathological features in the translated output is a major challenge.
• Recent advances in generative models, particularly diffusion models, have enabled impressive cross-modal medical image translation, but existing methods often prioritize structural preservation over pathology awareness.

Methodology

• PASTA employs a symmetric dual-arm architecture, consisting of a conditioner arm and a denoiser arm, connected through adaptive conditional modules.
• The conditioner arm generates task-specific multi-scale representations from the input MRI, which are used to condition the denoiser arm during the reverse diffusion process.
• Clinical data (age, gender, cognitive scores, genetic risk factors) are also integrated as additional conditioning signals.
• A cycle exchange consistency training strategy is introduced to promote informative feature extraction across the dual arms.
• PASTA utilizes a memory-efficient 2.5D volumetric generation strategy to produce high-quality 3D PET scans.
• Pathology-aware loss functions, including a learnable weight on disease-relevant regions, further enhance PASTA's ability to preserve clinically meaningful patterns.

Data & Experimental Setup

• Evaluated on two datasets: the Alzheimer's Disease Neuroimaging Initiative (ADNI) and an in-house clinical dataset from TUM Klinikum.
• Comprehensive quantitative and qualitative evaluations, including comparisons to state-of-the-art GAN and diffusion-based methods.
• Assessed Alzheimer's disease classification performance using the synthesized PET scans.
• Conducted fairness evaluation to ensure consistent performance across demographic groups.

Results

• Qualitatively, PASTA generates PET scans that closely match the ground truth, especially in preserving disease-specific pathological patterns like reduced glucose metabolism in Alzheimer's.
• Quantitatively, PASTA achieves the lowest reconstruction errors (MAE, MSE) and highest structural similarity (PSNR, SSIM) compared to all baselines.
• In Alzheimer's disease classification, PASTA's synthesized PET scans improve over MRI by over 4% in balanced accuracy, nearly reaching the performance of actual PET.
• PASTA also demonstrates superior pathology preservation within disease-relevant regions, as evidenced by localized metrics.
• Sensitivity analysis shows that cognitive scores (MMSE, ADAS-Cog) are the most influential clinical variables for PASTA's pathology-aware generation.
• Neurostat 3D-SSP analysis further confirms PASTA's ability to faithfully recover disease-specific metabolic patterns.

Interpretation

• PASTA's strong performance is attributed to its dual-arm architecture, multi-modal conditioning, and pathology-aware training objectives, which collectively enable effective learning of the complex structural and functional correlations between MRI and PET.
• The integration of relevant clinical data and disease-specific priors (MetaROIs) allows PASTA to capture complementary cues that enhance its sensitivity to abnormal metabolic changes.
• The cycle exchange consistency training strategy facilitates the extraction of richer, task-specific features, promoting the preservation of both structural integrity and pathological details during translation.

Limitations & Uncertainties

• The study lacks a formal clinical reading evaluation to quantitatively assess the diagnostic utility of PASTA's synthetic PET scans, which should be addressed in future work.
• While PASTA demonstrates significant improvements over existing methods, there is still room for further enhancing the fidelity and interpretability of the generated PET images.

What Comes Next

• Explore additional ways to improve the interpretability of PASTA's generation process, such as incorporating more explicit disease-specific constraints or integrating the model with downstream clinical decision support systems.
• Investigate the potential of PASTA's synthetic PET scans to augment real patient data for training more robust Alzheimer's disease classification models.
• Extend the PASTA framework to handle other neurodegenerative disorders and explore its broader applicability in cross-modal medical image translation tasks.