Purpose: This study we developed a deep learning (DL) dose prediction model for head- and-neck (H&N) cancer patients treated with Involved Nodal Radiation Therapy Using Artificial Intelligence-Based Radiomics (INRT-AIR), which eliminates the neck treatment by using radiomic-based metrics to focus radiation directly on the nodes themselves. Due to the relatively new procedure, we investigated the use of transfer learning to different pre-trained models and evaluated each of their performances against a baseline.
Methods: The dataset was acquired from 25 H&N patients treated with INRT-AIR containing CT, contours of the PTVs and 42 OARs of interest. The DL model architecture is a modified hierarchically densely connected U-net. We had previously trained and clinical implemented two versions of this model architecture on definitive and post-op H&N patients. For this study, we fine-tuned each of these models (definitive and post-op) onto the H&N INRT-AIR dataset, as well as training one model from scratch (baseline). We used the 3D dose distribution, dose difference error, and DVH to evaluate the different prediction dose.
Results: The results demonstrated discretely that the mean dose prediction error against the ground truth for the brain in the definitive dose are 0.11Gy, 1.93 Gy for the oral cavity, 2.79 Gy for the esophagus. The baseline and post-op dose had worse results, with a mean dose prediction error of 0.33Gy and 0.88Gy for the brain, 2.38Gy and 3.26Gy for the oral cavity, 1.89Gy and 1.61Gy for the esophagus. In addition, the overall definitive model has better coverage on D99 and D98 than original and post-op models on DVH.
Conclusion: We find that this new H&N INRT model that was fine-tuned for a definitive dose can achieve accurate and efficient outcomes. The proposed prediction model will be implemented in the clinical pipeline to improve plan quality and reduce planning times.
Funding Support, Disclosures, and Conflict of Interest: This project is supported by the National Institutes of Health (NIH)