Improving proton CT beyond iterative methods with a convolutional neural network

Relative stopping power (RSP) values of tissues in patients are needed to plan proton beam therapy accurately. Proton CT (pCT) is an alternative imaging method for obtaining more accurate RSP values than by using X-ray CT. This imaging modality gives mostly accurate RSP values but is blurred due to elastic multiple Coulomb scattering. To improve the blurriness of reconstructed pCT images, we have investigated a denoising convolutional neural network trained on known ground RSP values of a digital phantom. In our initial results, with the denoising network receiving pCT images reconstructed with an iterative method as input we observed improved spatial resolution and better RSP accuracy in the output images. The improved images had a higher peak signal-to-noise ratio (PSNR) and significantly improved structural similarity index measure (SSIM). More accurate RSP values with better spatial resolution will pave the way for more widespread adoption of pCT for proton treatment planning.