Dual-branch residual encoder-decoder convolutional neural network (DB-REDCNN): a computed tomography-integrated...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Positron emission tomography (PET) imaging involves radiation exposure due to radioactive tracers.
• Low-dose PET (LDPET) aims to reduce radiation dose while maintaining image quality comparable to full-dose PET (FDPET).
• Current deep learning methods struggle with denoising small lesions in LDPET.

Purpose of the Study:

• To develop a deep learning model that enhances LDPET image quality by incorporating structural information from CT.
• To improve the preservation of fine details and the imaging of small lesions in LDPET.
• To reduce radiation dose in PET imaging without compromising diagnostic quality.

Main Methods:

• Proposed a dual-branch residual encoder-decoder convolutional neural network (DB-REDCNN).
• The network utilizes paired CT images for structural priors to enhance PET reconstructions.
• Employed parallel branches for modality-specific feature extraction (PET and CT) with a fusion mechanism.

Main Results:

• DB-REDCNN demonstrated superior performance in quantitative metrics like root mean square error, PSNR, and structural similarity index.
• Achieved statistically significant improvements in edge sharpness (|K| values) compared to other methods (P<0.01).
• Markedly reduced SUV_max error, outperforming competing approaches in clinical indicator evaluation.

Conclusions:

• The DB-REDCNN model effectively synthesizes FDPET quality from LDPET by integrating CT structural priors.
• The approach enhances quantitative performance, lesion edge sharpness, and preserves crucial SUV information.
• Offers a clinically valuable solution for reducing radiation dose in PET imaging.