Deep-Learning PET Imaging for Shorter Scans and Accurate Quantification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Positron Emission Tomography (PET) systems require lengthy scan times, leading to patient discomfort, motion artifacts, and reduced image quality due to low counting statistics, which can compromise quantification accuracy.
Innovation Solution
Employ deep learning techniques to enhance PET image quality and preserve quantification accuracy by using a deep network model to transform low-quality images acquired with shortened scan times, and combine them with original images using an adaptive mixing algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If conventional PET imaging protocols are used, then image quality and quantification accuracy are maintained, but scan time becomes excessively long (10+ minutes)
Solution Approach 1:
The patent creates a synthetic PET image dataset by applying deep learning transformations to low-quality short-scan images. The neural network learns to generate high-quality images that replicate the appearance and quantitative characteristics of conventional long-scan images, effectively copying the quality without requiring the actual long scan time.
Solution Approach 2:
The patent changes the acquisition parameter from standard scan time to shortened scan time, then uses deep learning to compensate for the resulting image quality degradation. The system transforms images acquired with reduced scan time parameters into high-quality images through neural network processing.
2Productivity
If scan time is shortened, then patient comfort and productivity improve, but image quality deteriorates due to low counting statistics
Solution Approach 1:
The patent replaces the mechanical/data collection process of acquiring sufficient counts through prolonged scanning with a computational approach using deep learning. Instead of physically extending scan time to improve statistics, the system uses neural networks to synthesize high-quality images from limited data.
Solution Approach 2:
The system performs preliminary image acquisition with shortened scan time, then applies deep learning transformations to enhance quality. The neural network processes the low-quality images in advance to generate high-quality outputs, preparing the data for clinical use without requiring extended acquisition.
3Measurement precision
If deep learning models are applied to enhance image quality, then image quality improves, but computational complexity and processing time increase
Solution Approach 1:
The patent segments the image processing task into distinct stages: (1) acquiring PET data with shortened scan time, (2) applying deep learning transformations to enhance quality, and (3) combining enhanced images with original images using adaptive mixing. This segmentation allows each component to be optimized independently.
Solution Approach 2:
The system changes computational parameters by using adaptive mixing weights that dynamically adjust the contribution of enhanced versus original images. The mixing parameter is optimized to balance computational resources with image quality, allowing flexible adjustment based on specific clinical needs.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented method is provided for improving image quality with shortened acquisition time. The method comprises: determining an accelerated image acquisition parameter for imaging a subject using a medical imaging apparatus; acquiring, using the medical imaging apparatus, a medical image of the subject according to the accelerated image acquisition parameter; applying a deep network model to the medical image to generate a corresponding transformed medical image with improved quality; and combining the medical image and the corresponding transformed medial image using an adaptive mixing algorithm to generate output image.