Amyloid PET Translation Using Conditioned Diffusion From FDG PET
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current amyloid positron emission tomography (PET) imaging methods are limited in their ability to accurately synthesize amyloid PET images from Fluorodeoxyglucose (FDG) PET images, which are commonly used for glucose metabolism analysis, due to issues like mode collapse in generative adversarial network (GAN)-based architectures and the need for more advanced models to predict amyloid deposition patterns.
Innovation Solution
A conditioned diffusion-based generative model is employed to iteratively train FDG PET images using a weighted combination of sequential images, subject demographics, and disease status to synthesize high-quality amyloid PET images, leveraging a connection between FDG and amyloid PET imaging.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If GAN-based architectures are used for synthesizing amyloid PET images from FDG PET images, then image synthesis capability is improved, but mode collapse occurs resulting in limited variation and quality in generated images
Solution Approach 1:
The patent introduces a diffusion model as an intermediary between the input FDG PET image and the output amyloid PET image. Instead of using a direct GAN generator that suffers from mode collapse, the diffusion model provides a gradual, step-by-step transformation process through iterative denoising, mediating the image synthesis to avoid the pitfalls of direct adversarial generation while maintaining high image quality and diversity
Solution Approach 2:
The patent employs periodic action through the iterative nature of diffusion processes, where the image synthesis occurs in multiple discrete steps rather than a single transformation. Each step involves adding and removing noise in a controlled manner, allowing the model to progressively refine the generated image and avoid mode collapse by exploring multiple possible outcomes at each iteration
2Measurement precision
If amyloid PET imaging is used for directly assessing amyloid spatial distribution, then detection accuracy is improved, but cost and availability deteriorate
Solution Approach 1:
The patent creates a realistic copy of amyloid PET images by synthesizing them from FDG PET images using a diffusion model. Instead of requiring actual amyloid PET scans for every patient, the system generates high-quality synthetic copies that preserve the spatial distribution patterns and pathological features, making the information accessible through cheaper, more available FDG PET imaging
Solution Approach 2:
The patent makes FDG PET imaging multi-functional by enabling it to serve both its original purpose of assessing glucose metabolism and the additional function of predicting amyloid deposition patterns. This universality allows a single, widely available imaging modality to provide multiple diagnostic insights, eliminating the need for separate amyloid PET scans
3Measurement precision
If more FDG PET images are used for training to improve prediction accuracy, then model performance is improved, but training time and computational resources increase
Solution Approach 1:
The patent performs preliminary action by pre-processing the training data through careful selection and augmentation of FDG-amyloid PET image pairs before training begins. The diffusion model architecture is also pre-configured with appropriate noise schedules and hyperparameters, allowing the training process to converge faster with fewer iterations and reducing the overall training time while maintaining high prediction accuracy
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Amyloid deposition is considered a viable biomarker for Alzheimer's disease which is expensive and less used modality to study amyloid spatial distribution in brain. The present disclosure addresses problems of conventional approaches which exhibit trade-off between performance and mode collapse. This disclosure provides a system and method for subject-specific amyloid position emission tomography (PET) translation using conditioned diffusion-based generative model. The present disclosure discloses an architecture for synthesizing subject-specific amyloid images with a diffusion model. First effectiveness of a relationship between Fluorodeoxyglucose (FDG) and amyloid PET images is identified. Further, a framework is provided to synthesize Amyloid PET by utilizing its connection to FDG PET of same subject and cross-subject trends in amyloid deposition by incorporating age, gender and disease status information in learning process. In the other words, a diffusion model inspired image translation is provided to synthesize Amyloid PET from FDG PET and cross-subject amyloid deposition patterns.