Design of an Automated Process for Creating Breast Cancer Treatment Plans, with Artificial Intelligence: Development and Clinical Introduction of AI Models to Automate Segmentation and Planning for Breast Cancer Radiation Treatment Plans

Breast cancer is one of the most common cancer types in The Netherlands. Treatment of breast cancer often consists of (breast conserving) therapy, followed by post-operative radiotherapy. In order to perform radiotherapy, a treatment plan needs to be created, for which clinical target volumes (CTVs) and organs at risk (OARs) need to be identified and a dose distribution is calculated. Both steps of the treatment planning process involve iterative and manual actions. Besides the cumbersome nature of these steps, they are prone to the experience of the Radiotherapy Technologist (RTT) and Radiation Oncologist (RO), resulting in inter- and intra-observer variability. The goal of this design project, performed at the department of radiotherapy in the Catharina Hospital Eindhoven (CZE), is to develop and clinically introduce Artificial Intelligence (AI) models to automate the delineation of contours (auto-segmentation) and creation of the dose distribution (auto-planning).
For auto-segmentation, two AI models were developed, trained and evaluated, for both left- and right-sided breast cancer including the lymph nodes. The model training framework was provided by RaySearch Laboratories, including a 3D U-Net architecture. In total, 80 patients were included for training of both models, of which the contours were all visually inspected on abnormalities and corrected by two experienced RTTs and ROs, when needed. In a retrospective study, both models were tested for 15 patients: they showed to fulfill the predefined quantitative requirements for most of the cases. Therefore, a clinical pilot was performed in which the automatically generated contours were qualitatively scored by several RTTs and ROs. Besides, the time needed to automatically create the contours and perform corrections when needed was measured, too. A mean reduction in time of 42% was found for the OARs, while an even larger reduction of 59% was found for the CTVs. Furthermore, 92% of the contours were scored as clinically acceptable or useful for correction, indicating a high usability for clinical practice. For auto-planning, multiple models were developed, trained and validated for left-sided whole breast radiotherapy. During a previously performed project, two AI models had been trained and retrospectively validated for conventional breast irradiation (40.05 Gy in 15 fractions), using treatment plans of 90 patients. The first model was in-house developed, based on a 2D U-net architecture, whereas the second model was developed by RaySearch Laboratories, and based on a contextual Atlas Regression Forest (cARF). In this design project, both models were validated in a clinical pilot. Manually and automatically created plans were blindly scored by four experienced ROs, and the time to generate these plans was measured, too. Although there was a difference in preferences of the observers, 95% of the 2D U-Net plans were found to be clinically acceptable for all, which was the case in 90% of the manually generated and cARF plans. When only considering user-interaction time, both auto-planning methods showed time efficiency. Following the results of this study, a 3D U-Net was trained by Ray-Search, based on the same dataset, and was successfully commissioned for use at the department of radiotherapy in CZE. Hence, since May 2022, the model is used in clinic to generate treatment plans.
Besides, a 2D U-Net model was trained and retrospectively validated for fast-forward irradiation (2.6 Gy in 5 fractions). For this model, transfer learning was used, using the 2D U-Net for conventional irradiation as a starting point. This method proved to be promising, with good results while only using a dataset of 52 patients for training, and should be further investigated in the future for clinical use.
In conclusion, in this design project several AI models were successfully developed, trained and validated for delineation of contours and creation of dose distribution for breast cancer. While an autoplanning model was finally actually implemented in clinical practice, the auto-segmentation model showed promising results and will be clinically implemented in the near future.