Pelvic PET/MR attenuation correction in the image space using deep learning.

Area of Science:

• Medical Imaging
• Radiochemistry
• Artificial Intelligence

Background:

• Four-class Dixon-based attenuation correction (AC) in PET/MRI is error-prone due to bone.
• Existing five-class models that incorporate bone atlases are also known to be inaccurate.

Purpose of the Study:

• To develop a novel method for pelvic PET/MRI AC that directly predicts and corrects bone-related errors.
• To improve the accuracy of PET/MRI attenuation correction by addressing the limitations of current Dixon-based methods.

Main Methods:

• A convolutional neural network was trained to predict AC error maps using Dixon MR images and four-class AC µ-maps.
• The model was trained and validated using PET/MRI data from 22 patients and tested on 17 patients undergoing PSMA-based PET imaging.
• Quantitative analysis involved voxel- and lesion-based error metrics to assess performance.

Main Results:

• The proposed model reduced the median root mean squared percentage error from 12.1% and 8.6% to 6.2%.
• Median absolute percentage error in SUVmax for bone lesions improved from 20.0% and 7.0% to 3.8% compared to four- and five-class AC methods.
• The method effectively reduced voxel-based and SUVmax errors in bone lesions.

Conclusions:

• The novel convolutional neural network-based method significantly improves PET/MRI attenuation correction accuracy, particularly for bone lesions.
• This approach offers a more robust solution for attenuation correction in PET/MRI, outperforming existing Dixon-based models.