Related Experiment Video
Updated: Jul 27, 2025
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
9.3K
PET Image Denoising using a Deep-Learning Method for Extremely Obese Patients
Hui Liu1, Hamed Yousefi2, Niloufar Mirian2
1Department of Engineering Physics, Tsinghua University, and Key Laboratory of Particle & Radiation Imaging, Ministry of Education (Tsinghua University), Beijing, China, on leave from the Department of Radiology and Biomedical Imaging, Yale University, New Haven, CT, 06511, USA.
Summary
Deep learning noise reduction using U-Net improves clinical PET scan image quality in extremely obese patients. This method matches noise levels to lean subjects, preserving fine structures for consistent imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Clinical Positron Emission Tomography (PET) imaging quality suffers from high noise in extremely obese patients.
- This noise degradation impacts diagnostic accuracy and consistency.
- Standard noise reduction techniques may compromise image resolution.
Purpose of the Study:
- To reduce noise in PET images of extremely obese subjects to levels comparable to lean subjects.
- To ensure consistent imaging quality across different patient body types.
- To evaluate a deep learning-based noise reduction method for clinical PET scans.
Main Methods:
- A fully 3D patch-based U-Net deep learning model was employed for noise reduction.
- Two U-Nets (A and B) were trained on PET data from lean subjects at 40% and 10% count levels, respectively.
- The trained U-Nets were applied to PET images of 10 extremely obese subjects, with noise assessed via liver Normalized Standard Deviation (NSTD).
Main Results:
- U-Net A, trained on 40% count data, effectively reduced noise in obese patients' PET images (liver NSTD from 0.13±0.04 to 0.08±0.03, p=0.01).
- Post-denoising, noise levels in obese subjects matched those of lean subjects (0.08±0.03 vs. 0.08±0.02, p=0.74) while preserving fine structures.
- U-Net B (10% count data) resulted in over-smoothing and blurred image details.
Conclusions:
- A U-Net trained on lean subject data with matched count levels offers effective noise reduction for extremely obese patients in clinical PET scans.
- The method successfully maintained image resolution and consistency.
- Further clinical validation is recommended to confirm the utility of this deep learning approach.
