Feasibility of Deep Learning-Based PET/MR Attenuation Correction in the Pelvis Using Only Diagnostic MR Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Positron emission tomography/magnetic resonance (PET/MR) imaging requires accurate attenuation correction (AC) for quantitative analysis.
• Current AC methods often rely on dedicated MR sequences with limited diagnostic value, increasing scan times.
• Developing AC methods that utilize existing diagnostic MR images is crucial for efficiency and diagnostic utility.

Purpose of the Study:

• To evaluate the feasibility of deep learning-based attenuation correction (deepMRAC) using only diagnostically relevant MR images in the pelvis.
• To compare the accuracy of deepMRAC with conventional system-based MRAC for PET/MR imaging.
• To assess the impact of deepMRAC on quantitative PET values and lesion analysis.

Main Methods:

• A 3D deep convolutional neural network was trained using diagnostic T2 and T1 LAVA Flex MR images.
• The network generated a substitute CT (CTsub) from MR images for AC.
• CTsub was compared against reference CT (CTref) and used with system MRAC for PET/MR AC in test subjects.

Main Results:

• The deepMRAC approach achieved high Dice coefficients for soft tissue (0.98) and bone (0.79), with lower accuracy for bowel gas (0.49).
• Root mean square error for the whole PET image was significantly lower with deepMRAC (4.9%) compared to system MRAC (11.6%).
• Analysis of soft tissue lesions showed a narrower error distribution for maximum standardized uptake value with deepMRAC (-1.0% ± 1.3%) versus system MRAC (0.0% ± 6.4%).

Conclusions:

• Deep learning-based attenuation correction using only diagnostic MR images is feasible in the pelvis.
• DeepMRAC offers improved quantitative accuracy and reduced errors in PET/MR imaging compared to conventional methods.
• This approach has the potential to streamline PET/MR workflows by eliminating dedicated AC sequences.