Deep Auto-context Convolutional Neural Networks for Standard-Dose PET Image Estimation from Low-Dose PET/MRI

Area of Science:

• Medical imaging
• Radiology
• Artificial intelligence

Background:

• Positron Emission Tomography (PET) is crucial for diagnosing conditions like tumors and brain disorders.
• Standard-dose PET scans require radioactive tracers, posing radiation exposure risks.
• High-quality PET imaging is essential for accurate clinical diagnosis.

Purpose of the Study:

• To develop a deep learning model for estimating standard-dose PET (SPET) images from low-dose PET (LPET) and MRI data.
• To reduce radiation exposure while maintaining diagnostic image quality.
• To improve the efficiency of PET image reconstruction.

Main Methods:

• A deep learning architecture using convolutional neural networks (CNNs) was adapted for dual-channel input (LPET and T1-weighted MRI).
• An auto-context strategy was employed, integrating multiple CNN modules for iterative refinement of SPET images.
• The model learns an end-to-end mapping from input data to the desired SPET output.

Main Results:

• The proposed method achieved competitive estimation quality for PET images on real human brain PET/MRI data.
• The method demonstrated high efficiency, estimating a full SPET image in approximately 2 seconds.
• Performance was compared favorably against state-of-the-art methods.

Conclusions:

• The deep learning approach effectively estimates high-quality SPET images from LPET and MRI data.
• This technique significantly reduces patient radiation exposure and processing time.
• The method shows strong potential for integration into clinical practice.