Self-supervised deep learning for joint 3D low-dose PET/CT image denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Deep learning (DL) excels at denoising low-dose positron emission tomography (LDPET) and low-dose computed tomography (LDCT).
• Existing DL methods typically focus on single modalities and require extensive aligned low-dose/normal-dose sample pairs, which are difficult to acquire.
• Simultaneous denoising of LDPET and LDCT with a single network remains underexplored.

Purpose of the Study:

• To develop a self-supervised framework for joint LDPET/LDCT denoising.
• To overcome the limitation of requiring well-aligned low-dose/normal-dose sample pairs.
• To improve the robustness and reduce artifacts in denoised medical images.

Main Methods:

• A two-stage self-supervised training framework named MAsk-then-Cycle (MAC) was proposed.
• Stage 1: Masked autoencoder (MAE)-based pre-training for pixel-level constraint imposition.
• Stage 2: Cycle self-recombination (CSR) strategy for self-supervised denoising without aligned pairs, disentangling and recombining noise components.

Main Results:

• The proposed MAC framework achieved superior performance in joint LDPET/LDCT denoising compared to existing methods.
• The CSR strategy effectively enables denoising without requiring aligned sample pairs.
• MAE pre-training successfully imposed pixel-level constraints, mitigating artifacts common in self-supervised methods.

Conclusions:

• The MAC framework represents the first self-supervised method for joint LDPET/LDCT denoising.
• The approach is robust as it does not rely on prior assumptions about the data.
• This method offers a more flexible and effective solution for improving the quality of low-dose medical imaging.