Exhibit Hall | Forum 5
Purpose: Online adaptive proton therapy (APT) is a promising solution to mitigate the impact of inter-fractional changes during intensity modulated proton therapy (IMPT). It is however challenging to perform APT using conventional cone-beam CT (CBCT) systems, as those offer limited proton stopping power (RSP) accuracy and poor image contrast. In this work, we investigate the potential of dual-energy CBCT (DE-CBCT) to provide fast and accurate RSP estimation using deep convolutional neural networks (DCNN) to correct scatter in the projection domain for both energies in real time.
Methods: High and low energy projections and scatter distributions are simulated with Monte Carlo (MC) using 80 kVp and 120 kVp spectra for 10 head and neck patients for total of 7200 CBCT projections. Two different DCNN configurations (DE-CBCT₁, DE-CBCT₂) are trained to predict the scatter distributions from the raw DE-CBCT projections. A third DCNN is trained using the 120 kVp data only, serving as reference for single-energy CBCT (SE-CBCT). The accuracy of HU values reconstructed using each network configuration is evaluated against the reconstructed scatter free images. A numeric phantom derived from a patient not used for network training or testing is used as validation for APT. An IMPT plan optimized on the ground truth RSP values of the numeric phantom is recalculated on the RSP maps derived from scatter corrected DE-CBCT and SE-CBCT images to assess the performance of each approach.
Results: The mean absolute errors on HU for the low/high energies in the testing patient cohort are 26/16 and 23/8 using the DE-CBCT₁ and DE-CBCT₂ configuration respectively. 3%/3mm gamma pass rate of 98.2% and 97.2% are obtained for the IMPT dose distributions calculated with DE-CBCT₁ and DE-CBCT₂, compared to 94.7% using the SE-CBCT images.
Conclusion: DCNN-based scatter correction of DE-CBCT data enables fast and accurate RSP predictions suitable for online APT.
Cone-beam CT, Protons, Scatter
TH- External Beam- Particle/high LET therapy: Dual energy/spectral CT-based stopping power mapping