Transformation of PET raw data into images for event classification using convolutional neural networks.

Area of Science:

• Medical Imaging
• Machine Learning
• Nuclear Medicine

Background:

• Convolutional neural networks (CNNs) are used for pattern recognition in positron emission tomography (PET) imaging.
• Unprocessed PET coincidence data are typically in tabular format, posing challenges for direct CNN application.
• Existing methods require pre-processed PET data, limiting direct analysis of raw coincidence events.

Purpose of the Study:

• To develop a method for transforming tabular PET coincidence data into n-dimensional matrices.
• To prepare tabular PET data for classification using CNNs.
• To evaluate the performance of this data transformation method in classifying simulated PET events.

Main Methods:

• Developed a nonlinear transformation to convert tabular PET coincidence data into n-dimensional matrices.
• Utilized principal component analysis (PCA) for feature extraction and image creation from the transformed data.
• Applied the methodology to classify simulated PET coincidence events from NEMA IEC and XCAT phantoms using various neural network architectures (MLP, CNN).

Main Results:

• The developed method increased the number of features from 6 to 209.
• Achieved the highest precision (79.8%) among tested neural network architectures.
• Demonstrated minimal performance degradation when tested on data from a different phantom.

Conclusions:

• The proposed data transformation and feature engineering method effectively prepares tabular PET data for CNN-based classification.
• This approach enhances the utility of raw PET coincidence data for machine learning applications.
• The method shows robustness and improved classification precision in PET imaging analysis.