CNN Encoder-Decoder for Low-Dose PET Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for low-dose Positron Emission Tomography (PET) image reconstruction are either computationally expensive, vendor-specific, prone to over-smoothing, and fail to achieve satisfactory results at ultra-low doses, leading to radiation exposure risks for patients.
Innovation Solution
A deep learning method using a convolutional neural network with an encoder-decoder structure and symmetry concatenate connections is employed to generate standard-dose PET images from ultra-low-dose images, incorporating multi-contrast MRI data for improved image quality and dose reduction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative reconstruction algorithms are used to improve image quality for low-dose PET, then measurement precision is improved, but computing time increases substantially
Solution Approach 1:
The patent applies preliminary action by pre-training deep learning models on standard-dose PET images before deployment. The pre-trained models capture complex image priors and anatomical structures in advance, enabling rapid inference on low-dose images without requiring iterative reconstruction during actual scanning, thus resolving the contradiction between image quality and computing time
Solution Approach 2:
The patent substitutes traditional iterative reconstruction algorithms (mechanical/computational systems) with deep learning-based reconstruction networks. These neural networks replace the iterative optimization process with a single-pass inference that leverages learned priors from training data, achieving comparable or superior image quality with dramatically reduced computing time
2Measurement precision
If standard dose of radioactive tracer is injected to acquire high quality PET image, then measurement precision is improved, but radiation exposure risk increases
Solution Approach 1:
The patent introduces deep learning models as an intermediary between low-dose PET images and diagnostic-quality images. The models act as a mediator that transforms noisy low-dose inputs into high-quality outputs by learning the mapping from standard-dose training images, enabling dose reduction while maintaining diagnostic accuracy
Solution Approach 2:
The patent uses copying by training the deep learning models on standard-dose PET images that serve as reference copies. The models learn to replicate the quality characteristics of these standard-dose images when processing low-dose inputs, effectively creating synthetic copies of high-quality images from low-dose data without requiring actual high-dose exposure
3Measurement precision
If iterative reconstruction algorithms with regularization are used to reduce noise, then measurement precision is improved, but manufacturing precision deteriorates due to over-smoothing
Solution Approach 1:
The patent applies local quality by enabling the deep learning model to apply different processing strengths to different regions of the image. The network can preserve fine details in high-contrast regions while applying stronger denoising in homogeneous areas, avoiding the uniform over-smoothing effect of traditional regularization methods and maintaining both noise reduction and detail fidelity
4Measurement precision
If vendor-specific iterative methods are used to account for scanner geometry, then measurement precision is improved, but adaptability deteriorates
Solution Approach 1:
The patent applies universality by designing a vendor-agnostic deep learning framework that can be adapted to different scanner types and geometries through training data rather than hard-coded parameters. The same base model architecture can process images from various PET scanners by learning scanner-specific characteristics from training data, enabling one model to serve multiple vendor platforms without requiring vendor-specific algorithm development
Data Source
AI summary
A method of reducing radiation dose for radiology imaging modalities and nuclear medicine by using a convolutional network to generate a standard-dose nuclear medicine image from low-dose nuclear medicine image, where the network includes N convolution neural network (CNN) stages, where each stage includes M convolution layers having K×K kernels, where the network further includes an encoder-decoder structure having symmetry concatenate connections between corresponding stages, downsampling using pooling and upsampling using bilinear interpolation between the stages, where the network extracts multi-scale and high-level features from the low-dose image to simulate a high-dose image, and adding concatenate connections to the low-dose image to preserve local information and resolution of the high-dose image, the high-dose image includes a dose reduction factor (DRF) equal to 1 of a radio tracer in a patient, the low-dose PET image includes a DRF of at least 4 of the radio tracer in the patient.