MR Template-Based Individual Brain PET Volumes-of-Interest Generation Neither Using MR nor Using Spatial...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Precise quantification of mouse brain PET requires spatial normalization (SN) and template volumes-of-interest (VOIs).
• This method depends on corresponding MRI and SN, which are not always available for routine PET scans.
• Generating individual-brain-specific VOIs directly from PET images is needed.

Purpose of the Study:

• To develop a deep learning (DL) method for generating individual-brain-specific VOIs directly from PET images.
• To enable MR-less and SN-less PET image quantification.
• To apply this method to an Alzheimer's disease mouse model.

Main Methods:

• A deep convolutional neural network (CNN) model was trained using PET images as input and MR inverse-spatial-normalization (iSN)-based VOIs as labels.
• The method generated VOIs for cortex, hippocampus, striatum, thalamus, and cerebellum.
• The technique was validated on T2-weighted MRI and 18F FDG PET scans from 18 mice before and after treatment.

Main Results:

• The DL-based method achieved good VOI agreement (Dice similarity coefficient) and correlation of mean counts and SUVR.
• CNN-generated VOIs showed high concordance with ground-truth MR and MR template-based VOIs.
• Performance metrics were comparable to MR-based deep CNN VOI generation.

Conclusions:

• A novel MR-less and SN-less quantitative analysis method was established for generating individual brain space VOIs.
• This method utilizes MR template-based VOIs for PET image quantification.
• The approach offers a valuable tool for preclinical and clinical PET studies, particularly when MRI is unavailable.