Novel adversarial semantic structure deep learning for MRI-guided attenuation correction in brain PET/MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Quantitative PET/MR imaging relies on accurate synthetic CT (sCT) generation from MR images.
• Deep learning algorithms show promise for medical image analysis applications.

Purpose of the Study:

• To introduce a novel deep learning adversarial semantic structure (DL-AdvSS) algorithm for sCT generation.
• To improve MRI-guided attenuation correction in brain PET/MRI.

Main Methods:

• The DL-AdvSS algorithm uses an adversarial learning framework to ensure synthetic CTs align with structural features from real CT images.
• Evaluated against an atlas-based method (Atlas) and a commercial segmentation-based method (Segm) using clinical brain PET/CT and MR data from 40 patients.
• A two-fold cross-validation scheme was employed.

Main Results:

• DL-AdvSS and Atlas methods achieved similar high accuracy in cortical bone extraction (Dice coefficient ~0.77-0.78) and CT value estimation (mean error < -11 HU).
• The Segm approach showed significantly lower accuracy (mean error -1025 HU).
• PET quantitative analysis revealed DL-AdvSS and Atlas methods had SUV bias < 4% in brain regions and < 2% in cortical bone, while Segm resulted in ~15% SUV underestimation in cortical bone.

Conclusions:

• The DL-AdvSS approach demonstrates competitive performance compared to the atlas-based technique.
• DL-AdvSS achieves clinically tolerable errors and outperforms the commercial segmentation approach for brain PET/MRI attenuation correction.