Artificial Intelligence-Assisted Full-Process Solution for Rectal Cancer Radiotherapy

Abstract

To develop an artificial intelligence based full-process solution for rectal cancer radiotherapy. An integrated treatment planning method is developed that can finish whole treatment planning process without further manual work. A deep-learning based method is used to predict dose distribution and target as well as other organs at risk which can be performed as optimization function. Treatment planning system (TPS) auto planning is involved to create clinical accepted treatment plan. 40 rectal cancer patients with the same prescription doses are enrolled in this study. For each patient, dose volume histogram (DVH) are generated and transferred to the TPS. A script will automatically generate a treatment plan using the TPS. The OAR from predicted method is compared with those from manual segmentation. Additionally, the dose distribution from automated plan (Auto-plan) is compared to the manually optimized plan (MO-plan) and the predicted plan (Pre-plan). The DICE of segmentation between prediction and manual is: CTV 0.86±0.07, PTV 0.90±0.02, Left Femoral Head 0.81±0.11, Right Femoral Head 0.82±0.14, Bladder 0.81±0.10. 31 patients’ plans are clinical accepted without further modification while other 9 plans need modification before clinical use. Automated planning results: planned target area V99 = 99.09±0.24%, V95 = 99.97±0.02%, Dmax = 53.7±0.2Gy, CI = 1.09±0.03, HI = 1.09±0.01; left femoral head Dmean = 17.8±3.2Gy; right femoral head Dmean = 17.2± 2.4 Gy; bladder Dmean = 32.5 ± 3.6 Gy. Manual planning results: planned target area V99 = 98.64±0.79%, V95 = 99.93±0.10%, Dmax = 53.6±0.3Gy, CI = 1.04±0.03, HI = 1.08±0.01; left femoral head Dmean = 22.5±3.7Gy; right femoral head Dmean = 22.0± 3.5 Gy; bladder Dmean = 33.2 ± 3.2 Gy. Compared to the MO-plan, the Auto-plan shows better target coverage but a little poorer conformity. A deep-learning based fully automatic solution for rectal cancer treatment is established. The planning process and efficiency is improved by reducing time of contour segmentation and planning. This method is an elementary trial of automatic treatment planning.