Author: Keiichi Jingu, Noriyuki Kadoya, Takafumi Komiyama, Takeru Nakajima, Hikaru Nemoto, Hiroshi Onishi, Masahide Saito, Ryota Tozuka 👨🔬
Affiliation: Department of Radiology, University of Yamanashi, Department of Radiation Oncology, Tohoku University Graduate School of Medicine, Department of Advanced Biomedical Imaging, University of Yamanashi 🌍
Purpose: We evaluated the accuracy of a new AI-based fully automated planning software in stereotactic body radiotherapy (SBRT) for early-stage lung cancer.
Methods: We collected data from 125 patients with early-stage lung cancer treated with SBRT (55 Gy/4 Fr). We trained a deep learning model for automatic contouring, optimization, and PSQA using RatoGuide (Airato Inc., Sendai, Japan) with 120 training cases. Automatic planning was performed using five test cases. Planning involved five steps: (1) definition of the target on treatment planning computed tomography (CT) images by a radiation oncologist, (2) automatic contouring of seven risk organs, (3) dose distribution prediction using the automatic contours created in step 2, (4) automatic optimization by dose mimicking, and (5) PSQA by prediction of the gamma passing rate (GPR) at 2%/2 mm of the plan created in step 3. The accuracy of the AI plan was compared with that of the manual plan using the Dice coefficient and dose-volume histogram (DVH) parameters. The automatic contours and AI plan were visually scored by a radiation oncologist using a scale from 1 (totally unacceptable) to 5 (acceptable without any modifications).
Results: The average Dice coefficient between automatic and manual contouring for the seven at-risk organs was 0.85. The mean values of four DVH parameters (PTV-D98% [Gy], PTV-Dmax [Gy], Lung-V20 [%], SpinalCord-Dmax [Gy]) were 53.76 ± 0.23, 68.77 ± 1.60, 4.24 ± 1.35, 12.08 ± 6.70 for AI-based planning and 53.27 ± 0.35, 72.47 ± 2.06, 3.85 ± 1.35, 9.46 ± 1.06 for manual planning. The average error in predicting the GPR at a threshold of 2%/2 mm was 0.74 ± 0.98%. The average visual score by a radiation oncologist was 3.4 for automatic contouring and 4.4 for automatic optimization.
Conclusion: RatoGuide can potentially perform automated SBRT planning for early-stage lung cancer.