Deep Learning–Assisted Optimization of Y-90 Bremsstrahlung SPECT Using Multi-Energy-Window Monte Carlo Simulations
Abstract
Purpose
Artificial intelligence-based denoising has shown much promise in improving signal-to-noise in medical imaging. Our goal was to optimize Y-90 bremsstrahlung SPECT imaging following transarterial radioembolization (TARE) using a deep learning-based denoising model trained on various energy windows.
Methods
Training/test data for the 3D UNet denoising model were generated by GATE Monte Carlo (MC) simulation of projection data corresponding to multiple Y-90 phantoms (two Jaszczak phantoms at 3 sphere-to-background activity levels and an ellipsoid phantom with 25 and 30mm inserts) imaged on a Siemens Symbia T2 equipped with a MEGP collimator. Ground-truth (reference) activity maps were analytically defined via uniform activity assignment to sphere and background regions using Python. Due to limited computer resources for running long MC simulations, we took the approach of exploiting projection data from four energy windows of the Y-90 bremsstrahlung energy spectrum, which can be generated during the same simulation run to expand the training data set. Testing data corresponded to the 55-250 keV window which we use in clinical patient studies. Resulting projections were reconstructed using OSEM (15 iterations, 8 subsets). The model used a 75%/17%/8% train/validation/test split. A quantitative analysis using contrast recovery (CRC) relative to truth and contrast-to-noise ratio (CNR) was carried out to compare performance between UNet denoising vs post-reconstruction Gaussian filtering.
Results
UNet denoising outperformed Gaussian filtering, achieving substantial improvements in CNR (up to 2×) and CRC (closer to unity). It generalized well to unseen geometries. CRC and CNR for 30mm sphere Jaszczak were 0.98 and 18.72 respectively and 0.84 and 9.16 with Gaussian filtering. For the ellipsoid phantom, UNet denoising again outperformed Gaussian filtering, achieving superior CRC (0.64–1.06 vs 0.39–0.61) and CNR (4.75–8.08 vs 4.10–6.02) for 25–30 mm spheres.
Conclusion
The U-Net model proved to be better in its denoising capability across multiple phantom geometries.