Related Experiment Video
Updated: Jul 27, 2025
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
6.9K
Comparative study of the quantitative accuracy of oncological PET imaging based on deep learning methods
Yiyi Hu1,2,3, Doudou Lv1,2,3, Shaojie Jian1,2,3
1Department of Nuclear Medicine, First Hospital of Shanxi Medical University, Shanxi Medical University, Taiyuan, China.
Quantitative Imaging in Medicine and Surgery
|June 7, 2023
Summary
Deep learning models like 3D Unet and P2P can significantly improve [18F] Fluorodeoxyglucose (FDG) PET/CT image quality by reducing scan times. The 3D Unet model demonstrated superior enhancement in contrast-noise ratio for tumor lesions, meeting clinical diagnostic needs.
Area of Science:
- Medical Imaging
- Radiology
- Artificial Intelligence in Medicine
Background:
- [18F] Fluorodeoxyglucose (FDG) PET/CT is crucial for tumor assessment.
- Reducing scan time and radioactive tracer dose are key challenges.
- Deep learning offers solutions for optimizing PET/CT protocols.
Purpose of the Study:
- To evaluate deep learning models for reconstructing full-dose [18F]FDG PET/CT images from low-dose acquisitions.
- To compare the performance of Convolutional Neural Network (CNN) and Generative Adversarial Network (GAN) architectures.
- To determine the optimal neural network for accelerated PET/CT imaging.
Main Methods:
- Retrospective analysis of 311 tumor patients undergoing [18F]FDG PET/CT.
- Simulated low-dose PET images using 15s and 30s of each 3-min bed acquisition.
- Applied 3D Unet (CNN) and P2P (GAN) to predict full-dose images from low-dose inputs.
- Compared image quality scores, noise levels, and quantitative tumor parameters.
Main Results:
- High consistency in image quality scores across all deep learning groups (Kappa=0.719).
- Both 3D Unet and P2P models reduced background noise and increased signal-to-noise ratio (SNR).
- 3D Unet significantly improved contrast-noise ratio (CNR) and maintained quantitative accuracy (SUVmean, SUVmax) compared to standard protocols.
Conclusions:
- Generative Adversarial Network (GAN) and Convolutional Neural Network (CNN) models effectively reduce noise and enhance image quality in [18F]FDG PET/CT.
- 3D Unet excels in improving tumor lesion CNR while preserving quantitative accuracy, making it suitable for clinical diagnosis with reduced scan times.
