Dual Flow Generative Computer Architecture for MRI to PET Modality Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for modality transfer from MRI to PET images using CNNs and GANs suffer from blurry outputs and mode collapse, which result in loss of diagnostic differences, and are computationally challenging.
Innovation Solution
A machine learning architecture comprising two networks joined by a statistical model that imposes a predefined statistical model family, allowing for constrained training and error propagation to prevent averaging and mode collapse, while incorporating conditional probability for improved dimensionality control and integration of side information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If GANs are used for modality transfer from MRI to PET images, then sharper images are produced, but mode collapse occurs producing one of a limited set of output images for a much larger set of input images
Solution Approach 1:
The patent implements a feedback mechanism where the statistical model continuously monitors and evaluates the output images from the GAN, comparing them against the predefined statistical model family. This feedback loop detects mode collapse and adjusts the generator's parameters accordingly, ensuring diverse and accurate PET image generation while maintaining sharpness.
Solution Approach 2:
The patent changes the parameters of the statistical model family to better match the distribution of actual PET images. By adjusting these statistical parameters during training, the system can generate a more diverse set of output images that accurately represent the underlying disease states, preventing mode collapse while maintaining image quality.
2Productivity
If CNNs are used for modality transfer from MRI to PET images, then computational efficiency is improved, but blurry output images are produced when compared to GANs
Solution Approach 1:
The patent merges the advantages of both CNNs and GANs by combining the computational efficiency of CNNs with the image quality generation capabilities of GANs. The hybrid architecture uses CNNs for efficient feature extraction and processing, while GANs generate sharp and diverse PET images, achieving both computational efficiency and high image quality.
3Manufacturing precision
If GANs are used to characterize probability distribution from training images, then sharper images are produced, but training becomes computationally difficult
Solution Approach 1:
The patent segments the training process into distinct stages: first training the statistical model with a subset of data to establish the probability distribution, then using this pre-trained model to guide the GAN training. This segmentation reduces the computational difficulty by breaking down the complex training task into manageable phases, while still achieving sharp image generation.
Data Source
AI summary
A machine learning architecture employs two machine learning networks that are joined by a statistical model allowing the imposition of a predetermined statistical model family into a learning process in which the networks translate between and data types. For example, the statistical model may enforce a Gaussian conditional probability between the latent variables in the translation process. In one application, MRI images may be translated into PET images with reduced mode collapse, blurring, or other “averaging” type behaviors.