Noise reduction with cross-tracer and cross-protocol deep transfer learning for low-dose PET

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Nuclear Medicine

Background:

• Deep learning methods have shown promise in reducing noise in low-dose fluorodeoxyglucose (FDG) positron emission tomography (PET).
• Training deep learning models often requires large, specific datasets, which may not be available for all PET tracers or scanning protocols.
• Transfer learning offers a potential solution by leveraging existing models trained on abundant data.

Purpose of the Study:

• To investigate the feasibility of noise reduction for PET tracers with limited training data using deep transfer learning.
• To evaluate the effectiveness of a deep transfer learning approach utilizing pre-trained networks from FDG datasets for cross-tracer and cross-protocol applications.
• To compare the performance of different U-Net configurations for denoising PET images.

Main Methods:

• A fully 3D patch-based U-Net architecture was employed for noise reduction.
• Deep transfer learning was implemented using pre-trained U-Nets from 18F-FDG datasets (single-bed and whole-body).
• These pre-trained models were fine-tuned on smaller datasets of 18F-fluoromisonidazole (18F-FMISO) and 68Ga-DOTATATE, with full-dose images serving as ground truth.

Main Results:

• All evaluated U-Net models produced denoised images of comparable quality across different tracers.
• No significant differences in normalized root-mean-square error (NRMSE) or signal-to-noise ratio (SNR) were observed between models fine-tuned with limited data and those trained extensively.
• Fine-tuning a single-bed FDG pre-trained U-Net with whole-body DOTATATE data yielded the most consistent results, significantly reducing NRMSE and ROI bias while improving SNR.

Conclusions:

• Deep transfer learning is a feasible strategy for noise reduction in low-dose PET imaging, particularly for tracers with limited datasets.
• Fine-tuning pre-trained deep learning models allows for effective noise reduction across different tracers and scanning protocols.
• This approach enhances image quality and quantitative accuracy in PET imaging without requiring extensive tracer-specific training data.