Enhanced Attenuation Correction in Hybrid PET/MR Imaging Using a Deep-Learning-Based Continuous μ-Map Generation Framework

Motivation: Accurate attenuation correction (AC) in PET/MR imaging faces challenges due to MR-based AC limitations, especially in segmenting bones and air-filled cavities, impacting PET quantification. Goal(s): Develop a deep-learning-based MRAC framework with a pseudo-CT model and continuous μ-maps to improve PET quantification without specialized MR sequences. Approach: The framework includes MR-to-CT generation, registration, and segmentation modules. Trained on 300 subjects, it was evaluated on 17 clinical cases, comparing four-tissue and five-tissue MRAC models to CTAC. Results: The five-tissue continuous μ-map (MRAC4) showed high PET quantification accuracy, closely aligning with CTAC, and reduced dependency on specialized MR sequences. Impact: The proposed framework enhances PET/MR accuracy, benefiting oncology and neurology diagnostics.