Machine Learning Reconstruction for Low-Noise SPECT Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing SPECT imaging systems suffer from low spatial resolution and high noise due to the use of collimators, limiting their application in quantitative studies and clinical diagnostics.
Innovation Solution
A network of machine learning models, comprising a first model with fully connected layers and a second model with convolutional layers, processes SPECT and CT data to optimize image reconstruction by filtering noise and compensating for attenuation and resolution blur, leveraging an attenuation map to improve image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If collimators are used in SPECT imaging systems, then radiation safety and detector protection are improved, but spatial resolution deteriorates and noise increases
Solution Approach 1:
The patent replaces the mechanical collimator system with a machine learning-based reconstruction system. Instead of using physical collimators to restrict radiation paths, the invention uses neural networks to process SPECT data and generate high-resolution images, thereby eliminating the spatial resolution limitations imposed by mechanical collimators while maintaining radiation safety benefits
Solution Approach 2:
The patent changes the fundamental parameter of image reconstruction from traditional iterative algorithms to machine learning models. By training neural networks on simulated and real SPECT data, the system learns to map raw detector counts to high-resolution images, transforming the reconstruction process from physics-constrained to data-driven while improving spatial resolution
2Reliability
If collimators are used in SPECT imaging systems, then radiation safety is improved, but noise increases
Solution Approach 1:
The patent substitutes the mechanical collimator with an AI-based noise filtering system. The machine learning models are trained to distinguish between signal and noise patterns in SPECT data, enabling the system to reduce noise while maintaining diagnostic quality, thereby eliminating the noise penalty associated with collimator use
Solution Approach 2:
The patent converts the harmful noise introduced by collimators into a beneficial training signal. By incorporating noisy SPECT data into the machine learning training process along with ground truth images, the system learns to denoise and reconstruct high-quality images from low-quality inputs, turning the noise problem into an opportunity to improve image quality through learned transformations
3Device complexity
If traditional image reconstruction methods are used, then computational simplicity is maintained, but image quality and diagnostic accuracy deteriorate
Solution Approach 1:
The patent uses machine learning models to copy and transform raw SPECT data into high-resolution images. The neural networks are trained on pairs of raw data and corresponding ground truth images, learning to replicate the quality and detail of reference images from noisy inputs, thereby achieving high image quality without complex iterative reconstruction algorithms
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer-storage media, for machine learning image reconstruction. In some implementations, first input data representing the image of the one or more internal structures generated using a first imaging device is provided as input to a first machine learning model having one or more fully-connected layers. First output data generated by the first machine learning model is obtained and the first output data together with second input data representing a second image of the one or more internal structures generated using a second imaging device is provided to a second machine learning model having one or more convolutional layers. Second output data generated by the second machine learning model is obtained and used to generate rendering data that, when processed by a computing device, causes the computing device to output a reconstructed image.