A fast fully connected neural network-based method for random and scattered coincidence correction in polyhedral brain PET
Yuxuan Zhong, Senhao Wang, Xiulian ChenAbstract
Background
Scattered and random coincidences in brain PET imaging introduce significant noise and degrade resolution. Current correction algorithms for these effects are complex and computationally intensive. Fully connected neural network (FCNN) offers strong feature learning capabilities, versatility, and flexibility. Given that the coincidence data in brain PET imaging has relatively few features, FCNNs are well-suited to rapidly establish the relationship between coincidence data information and coincidence event types.
Methods
Eleven digital brain models were simulated using the Geant4 application for tomographic emission (GATE) [1]. Features extracted from the resulting data were used as inputs, with coincidence type as the target. FCNN was trained on this dataset. Quantitative evaluation was performed using contrast recovery coefficient (CRC), signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) metrics.
Results
FCNN enables rapid filtering of coincidence data outside the brain region and efficiently corrects both random and scattered coincidences within the brain region. After valid coincidence discrimination, the valid coincidence rate remained at 95.60%, while the scatter fraction decreased from 40.60% to 29.01% and the random fraction dropped from 44.65% to 33.85%. Quantitative evaluation demonstrates that the random and scatter coincidence correction (RSC) method significantly enhances image quality when combined with attenuation-corrected (AC), improving the CRC from 0.946 to 0.969, SNR from 10.24 to 19.13, and CNR from 0.402 to 0.496. Conversely, under non-attenuation-corrected (NAC) conditions, RSC improves CRC (0.945 to 0.971) but reduces SNR (11.53 to 9.73) and CNR (0.569 to 0.521), indicating a trade-off between signal purity and statistical noise.
Conclusion
FCNN effectively reconstructs the true activity distribution of the radioactive source and enhances image contrast, demonstrating promise for clinical real-time brain PET imaging.