Machine learning-based approach reveals essential features for simplified TSPO PET quantification in ischemic stroke...

Area of Science:

• Nuclear Medicine
• Radiochemistry
• Machine Learning

Background:

• Neuroinflammation assessment post-ischemic stroke aids treatment selection.
• Translocator protein PET (TSPO PET) visualizes neuroinflammation in vivo.
• Standard TSPO PET quantification is complex, requiring long scans and arterial sampling.

Purpose of the Study:

• To develop a simplified TSPO PET quantification method using machine learning.
• To identify essential data for accurate V_T estimation.
• To reduce the complexity of routine TSPO PET analysis.

Main Methods:

• Analyzed [¹⁸F]GE-180 PET and MRI data from 18 ischemic stroke patients.
• Used Logan plot with dynamic PET and image-derived input function (IDIF) for V_T calculation.
• Trained a random forest regressor with late PET frames, ASL, MRI data, and plasma concentrations.

Main Results:

• Machine learning model predicted V_T with 87.8% accuracy.
• Late PET frames and plasma concentrations were key predictors.
• A simplified method (late PET frame / plasma concentration) showed high accuracy, especially using the 70-80 min frame and 30 min sample.

Conclusions:

• Accurate TSPO PET quantification is achievable with simplified methods.
• A single late PET frame divided by a late blood sample concentration provides reliable V_T estimates.
• This approach simplifies routine clinical application of TSPO PET.