Room: Exhibit Hall | Forum 7
Purpose: To reliably estimate the proton stopping-power-ratio (SPR) with dual-energy CT imaging by employing a convolutional neural network (CNN) approach.
Methods: Voxelized computational phantoms of various sizes (i.e., 55-135 kg male with 5 kg increment) were obtained from National Cancer Institute (NCI). 125 slices per phantom were taken from the pelvic to the breast region. Each slice was forward projected with 80 kVp and 150 kVp/Sn spectrum information to created artificial CT images. A standard beam hardening correction based on water was applied to the projection prior to reconstructing the image. The reference material information was used to create the ground-truth SPR map of each slice using the Bethe-Bloch equation. A U-net structure taking the DECT images as the input and SPR map as the output will be trained with 14 male phantoms and tested with the other 3 phantoms. To evaluate our deep-learning approach, a widely referred conventional parametric model will be calibrated with the same NCI phantoms and compared.
Results: We have refined the images and constructed a U-net model for SPR map prediction. By utilizing the wealth of information obtained from various size phantoms, we expect that the deep learning model will reduce the uncertainty caused by imaging artifacts, which is the dominant source of uncertainty in SPR estimation. Moreover, impact from modelling inaccuracy and random noise will also likely be reduced. Overall we anticipate that our deep learning model will largely outperform the conventional pixel-wise model.
Conclusion: The deep learning approach have the potential to reduce the proton range uncertainty. Once the model's performance and generalizability is confirmed, our future goal will be to translate the learning obtained from simulated images to the real patient images.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the Cancer Prevention and Research Institute of Texas (CPRIT) grant (RP160661).