Related Experiment Video
Updated: Jan 15, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
LeqMod: Adaptable Lesion-Quantification-Consistent Modulation for Deep Learning Low-Count PET Image Denoising
IEEE Transactions on Medical Imaging
|October 6, 2025
Summary
A new deep learning strategy, LeqMod, enhances positron emission tomography (PET) image denoising by improving lesion visibility and quantification accuracy. This method reduces radiation exposure while preserving critical details in medical imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Deep learning-based positron emission tomography (PET) image denoising aims to reduce radiation exposure and scan times.
- Current methods often blur important details, negatively impacting lesion quantification accuracy.
- Accurate quantification is crucial for effective diagnosis and treatment monitoring in PET imaging.
Purpose of the Study:
- To introduce a novel strategy, LeqMod (lesion-perceived and quantification-consistent modulation), for enhanced PET image denoising.
- To improve lesion visibility and quantification consistency in denoised PET images.
- To develop a plug-and-play method adaptable to various deep learning architectures without increasing inference computational load.
Main Methods:
- The LeqMod strategy integrates downstream lesion quantification analysis as auxiliary tools during model training.
- It comprises two components: LeMod (lesion-perceived modulation) and QuMod (multiscale quantification-consistent modulation).
- LeMod uses differential sampling weights and loss criteria for lesion-present versus lesion-absent samples, guided by an auxiliary segmentation network.
- QuMod focuses on improving the accuracy of standardized uptake value (SUVmean and SUVmax) across multiple scales and sub-regions.
Main Results:
- Experiments on large, multi-center, multi-vendor PET datasets demonstrated LeqMod's effectiveness across different denoising frameworks.
- Integration of LeqMod reduced average lesion SUVmax bias by 5.92% compared to non-LeqMod frameworks.
- LeqMod improved the average peak signal-to-noise ratio (PSNR) by 0.36 across participating sites.
Conclusions:
- The proposed LeqMod strategy significantly enhances PET image denoising by preserving crucial details and improving quantification accuracy.
- LeqMod offers a versatile and computationally efficient solution for improving the quality and diagnostic utility of low-count PET images.
- This approach holds promise for reducing radiation dose and scan times while maintaining high diagnostic performance in clinical PET imaging.
