Deep learning applications for quantitative and qualitative PET in PET/MR: technical and clinical unmet needs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Quantitative and qualitative Positron Emission Tomography (PET) in hybrid PET/Magnetic Resonance Imaging (PET/MR) systems offers significant diagnostic potential.
• Deep learning (DL) presents promising avenues for improving PET/MR image quality and quantitative accuracy.
• However, several technical and clinical challenges hinder the widespread adoption of DL in PET/MR applications.

Purpose of the Study:

• To provide a comprehensive overview of the unmet needs and challenges in applying DL for quantitative and qualitative PET in PET/MR.
• To highlight potential solutions and research opportunities for overcoming these limitations.
• To facilitate the clinical translation of DL techniques in PET/MR imaging.

Main Methods:

• Review of current DL applications in PET/MR focusing on attenuation correction, image enhancement, motion correction, kinetic modeling, and data generation.
• Identification of data scarcity and technical limitations in DL-based attenuation correction (DLAC), particularly for pediatric and lung imaging.
• Evaluation of DL for image enhancement, motion correction, and kinetic modeling, noting challenges with clinical validation and data availability.

Main Results:

• DL-based attenuation correction (DLAC) is underexplored for pediatric and lung PET/MR due to data shortages.
• DL image enhancement shows potential but requires further clinical validation across radiotracers and motion-prone areas.
• Acquiring paired motion-corrupted and corrected PET/MR data for DL motion correction training is a significant hurdle.
• DL can mitigate limitations of dynamic PET, such as long scan times and motion artifacts.
• Monte-Carlo simulations offer a solution for generating large datasets to overcome clinical data scarcity.

Conclusions:

• Significant technical and clinical challenges remain for DL in PET/MR, including data availability and validation.
• Further research into DLAC, motion correction, and robust data generation strategies is crucial.
• Addressing these unmet needs will accelerate the clinical translation of DL-powered PET/MR solutions.