Deep Learning Low-Dose PET Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Low-dose PET imaging results in higher relative noise and loss of finer details in reconstructed images, leading to potential misdiagnoses due to reduced image quality.
Innovation Solution
A deep learning-based method using an encoder-decoder residual network with concatenate skip connections is employed to reconstruct low-dose PET images, applying a neural network to estimated full-dose SUV images and incorporating a loss function with mean square error and additional smoothing components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If a lower dose of radiopharmaceutical is used, then radiation exposure to patients and technicians is reduced, but image quality deteriorates with higher relative noise and loss of finer details
Solution Approach 1:
The patent introduces an intermediary deep learning model that mediates between the low-dose PET image and the full-dose PET image. The model learns the mapping relationship from low-dose to full-dose images and uses this learned relationship to generate high-quality full-dose images from low-dose inputs, thereby eliminating the need to actually administer high doses of radiopharmaceutical while still obtaining high-quality images
Solution Approach 2:
The patent changes the parameter of radiopharmaceutical dosage from high dose to low dose, and compensates for the resulting image quality degradation through post-processing. The deep learning model adjusts image parameters (noise levels, contrast, detail preservation) to transform the low-dose image characteristics into full-dose image characteristics
2Manufacturing precision
If post-acquisition image processing techniques are applied, then image quality can be improved, but processing time and computational complexity increase
Solution Approach 1:
The patent replaces traditional mechanical/image-processing methods (iterative reconstruction algorithms, regularization techniques) with a deep learning-based neural network. The neural network performs image enhancement through learned patterns rather than through complex iterative calculations, significantly reducing processing time while maintaining or improving image quality
3Reliability
If iterative reconstruction algorithms are used, then image reconstruction can be performed, but the process is complicated and slow
Solution Approach 1:
The patent substitutes complex iterative reconstruction algorithms with a pre-trained deep learning model. Instead of performing complicated iterative calculations during image reconstruction, the system uses the learned mapping from the neural network to directly generate high-quality images, simplifying the reconstruction process while maintaining reliability
Solution Approach 2:
The patent performs the complex algorithmic work in advance by pre-training the deep learning model on pairs of low-dose and full-dose images. This preliminary training phase captures the reconstruction knowledge, which is then applied quickly during actual image processing without requiring complex iterative algorithms at runtime
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Emission imaging data are reconstructed to generate a low dose reconstructed image. Standardized uptake value (SUV) conversion (30) is applied to convert the low dose reconstructed image to a low dose SUV image. A neural network (46, 48) is applied to the low dose SUV image to generate an estimated full dose SUV image. Prior to applying the neural network the low dose reconstructed image or the low dose SUV image is filtered using a low pass filter (32). The neural network is trained on a set of training low dose SUV images and corresponding training full dose SUV images to transform the training low dose SUV images to match the corresponding training full dose SUV images, using a loss function having a mean square error loss component (34) and a loss component (36) that penalizes loss of image texture and/or a loss component (38) that promotes edge preservation.