deepPERFECT: Novel Deep Learning CT Synthesis Method for Expeditious Pancreatic Cancer Radiotherapy

BACKGROUND: Pancreatic cancer is a devastating disease with more than 60,000 new cases each year and less than 10 percent 5-year overall survival. Radiation therapy (RT) is an effective treatment for Locally advanced pancreatic cancer (LAPC). The current clinical RT workflow, however, is lengthy and involves separate image acquisition for diagnostic CT (dCT) and planning CT (pCT) which imposes a huge burden on patients and their caretakers. Moreover, studies have shown a reduction in mortality rate from expeditious radiotherapy treatment course. Although, in theory, dCT can be used for RT planning, the differences in the image acquisition setup and patient’s body demand a new scan to be acquired. PURPOSE: To address this issue, we are presenting deepPERFECT: deep learning-based Planning External-beam Radiotherapy Free from Explicit simCT, that adapts the shape of the patient body on dCT to the treatment delivery setup. Our method expedites the treatment course by allowing the design of the initial RT planning before the pCT acquisition. Thus, the physicians can evaluate the potential RT prognosis ahead of time, verify the plan on the treatment day-one CT and apply any online adaptation if needed. METHODS: We used the data from 25 pancreatic cancer patients undergone stereotactic body radiation therapy. Each patient had a pair of dCT and pCT as part of their treatment course. The model was trained on 15 cases and tested on the remaining ten cases. We evaluated the performance of four different deep-learning architectures for this task. The synthesized CT (sCT) and regions of interest (ROIs) were compared with ground truth (pCT) using Dice similarity coefficient (DSC) and Hausdorff distance (HD). Finally, we evaluated the RT plan dose distribution for four scenarios. RESULTS: We found that the three-dimensional Generative Adversarial Network (GAN) model trained on large patches has the best performance. Using the in-place deformed image identity loss enhanced the performance of the deepPERFECT in predicting the body shape. The synthesized deformation fields and CT scans were evaluated using multiple figures of merit. The average DSC and HD for body contours were 0.93, and 4.6 mm. Additionally, we evaluated the quality of clinical-grade radiotherapy plans designed using the synthesized CT, by comparing the dosimetric indices measured on synthesized CT and ground truth. We found no statistically significant difference between the synthesized CT plans and the ground truth. CONCLUSIONS: We showed that deepPERFECT predicts the shape of the patient body on pCT using the dCT scan with good performance. We believe employing deepPERFECT shortens the current lengthy clinical workflow by at least one week and improves the effectiveness of treatment and the quality of life of pancreatic cancer patients.