Semisupervised Tripled Dictionary Learning for Standard-Dose PET Image Prediction Using Low-Dose PET and Multimodal...

Area of Science:

• Medical Imaging
• Radiology
• Machine Learning

Background:

• Positron emission tomography (PET) imaging often requires high tracer doses for quality.
• Low-dose PET (L-PET) imaging reduces radiation exposure but yields lower image quality.
• Magnetic resonance imaging (MRI) provides complementary anatomical information.

Purpose of the Study:

• To develop a method for predicting standard-dose PET (S-PET) images from L-PET and MRI data.
• To improve PET image quality using low-dose tracer injections.
• To leverage incomplete datasets for enhanced prediction performance.

Main Methods:

• Utilized patch-based sparse representation (SR) for initial dictionary construction.
• Developed a semisupervised tripled dictionary learning (SSTDL) method.
• Incorporated both complete and incomplete multimodal training samples (MRI, L-PET, S-PET).

Main Results:

• The proposed SSTDL method demonstrated superior performance compared to SR and other baseline methods.
• Validation was conducted on a real human brain dataset with 18 subjects.
• The method effectively predicts S-PET images from L-PET and MRI.

Conclusions:

• The developed semisupervised method significantly improves PET image quality from low-dose injections.
• This approach offers a promising solution for clinical applications by reducing radiation risk.
• The SSTDL method effectively utilizes incomplete datasets to enhance prediction accuracy.