Author: Pradeep Bhetwal, Wookjin Choi, Adam Dicker, Rupesh Ghimire, Yingcui Jia, Lauren Nkwonta, Yevgeniy Vinogradskiy, Wentao Wang, Maria Werner-Wasik 👨🔬
Affiliation: Thomas Jefferson University 🌍
Purpose: Multi-disciplinary clinics are becoming standard of care for patients with lung cancer treated with SBRT. To improve clinical decision support in a multi-disciplinary clinic, it would be beneficial to have information available on doses to organs at risk (OAR). However, doses to OARs are not available at the time of a multi-disciplinary clinic as the treatment plan has not been made yet. The purpose of this work was to develop an AI-based dose prediction algorithm using only a diagnostic PET/CT scan for lung SBRT.
Methods: 112 patients in different stages of lung cancer treated with SBRT were analyzed. Doses from the planning CT images were rigidly registered to the PET/CT images. OAR contours were auto-segmented using MIM ProtegeAI on planning CT, transferred to PET/CT, and converted into binary OAR masks. Two types of AI model architectures, UNET and Attention UNET, were trained on retrospective PET/CT or PET only or CT only inputs along with the OAR masks to predict dose distribution. Data were randomly split into 80% training and 20% for validation. Model performance was evaluated comparing the actual clinical and UNET predicted values differences in mean heart dose (MHDD(Gy)) or Lung V20Gy(%). The values close to 0 indicate good agreement.
Results: Median MHDD (Gy) for UNET across inputs (PET/CT; CT; PET) were (-0.04±2.79Gy;0.11±3.02Gy;-0.27±2.76Gy) respectively, and for Attention UNET were (0.06±1.54Gy;0.04±2 .15Gy; 0±2.15Gy) respectively. For Lung V20Gy (%), UNET values were (3.09±2.93%;3.61±2.98%;2.94±3.63%) respectively, and Attention UNET values were (1.74±3.09%; 2.26±3.08%; 2.09±3.65%) respectively.
Conclusion: The data showed that AI, PET-CT-based models can predict doses to OARs for patients being treated with SBRT with good accuracy. The PET/CT inputs provided the best predictions, closely followed by CT only predictions. Providing OAR doses to clinicians in a multi-disciplinary setting prior to the treatment plan being created can improve clinical decision-making.