Author: Daisuke Kawahara, Takaaki Matsuura, Yuji Murakami, Ryunosuke Yanagida 👨🔬
Affiliation: Hiroshima High-Precision Radiotherapy Cancer Center, Department of Radiation Oncology, Graduate School of Biomedical and Health Sciences, Hiroshima University, Department of Radiation Oncology, Graduate School of Biomedical and Health Sciences, Hiroshima University, Hiroshima 🌍
Purpose: In recent years, automation in radiation therapy planning using AI has gained significant attention to reduce the workload of treatment planners. Adaptive Radiation Therapy (ART), as a new form of radiation therapy, has further highlighted the need for rapid and efficient plan creation. This study is the first to establish a fully automated IMRT planning workflow that does not require optimization, enabling the generation of clinically deliverable plans in real-time.
Methods: We developed a two-step AI-based framework for fully automated IMRT planning. First, an improved ResCascade U-Net predicted dose distributions for individual beams from CT and structure data (1st dose prediction). Next, the ResVit network, a hybrid model combining CNNs and Transformers, generated fluence maps based on the predicted dose distributions. These fluence maps were converted into a format compatible with Eclipse, enabling clinically deliverable plans without further optimization (2nd dose prediction). Using data from 50 prostate cancer patients, the framework was trained and validated. Dose-Volume Histogram (DVH) metrics confirmed its accuracy for the target and organs at risk (OARs).
Results: For the test 5 cases, the mean relative errors of the DVH metrics of the PTV-rectum for the 1st and 2nd dose predictions were as follows: D98% was 0.4±0.1% and 2.1±0.8%, D95% was 0.3±0.2% and 1.7±0.8%, D50% was 0.2±0.1% and 0.1±0.2%, and D2% was 0.2±0.1% and 2.2±0.3%. For the OARs, the bladder had a Dmax of 0.3±0.2% and 3.9±1%, and the Dmean of 0.7±0.4% and 2.4±1% for the 1st and 2nd dose predictions, respectively. The rectum showed a Dmax of 0.6±0.3% and 0.9±0.4%, with a Dmean of 0.5±0.3% and 3.0±2%.
Conclusion:
This study successfully achieved complete automation of IMRT planning using AI, enabling instantaneous plan creation. This approach not only reduces the workload for planners but also has the potential to significantly advance the widespread adoption of ART.