Abstract
BACKGROUND: Synthetic computed tomography (sCT) images generated using deep learning (DL) methods enable the use of on-board CBCT imaging systems for online adaptive proton therapy workflows. However, DL models are susceptible to data drifts, such as changes in the quality of the CBCT images due to software upgrades. PURPOSE: This study aims to assess the effectiveness of transfer learning strategies in addressing changes in the input image quality and to evaluate the sustainability of potential sCT-dependent workflows following CBCT software upgrades. METHOD: Transfer learning strategies were utilized to re-train two existing DL-based sCT models (DCNN and cycleGAN). A dataset comprised 69 head and neck (HNC) patients with paired CBCT-CT images acquired after an image reconstruction software upgrade were used for this study. 60 patients were used for training and validation, and the remaining 9 were reserved for testing. To assess the efficacy of transfer learning strategies, several transfer learning models (TL-models) were trained using various subsets of data, ranging from 5 to 40 image pairs. Additionally, a post-upgrade sCT (New(PU)) model was trained utilizing the complete set of 60 patients to benchmark the TL-models to a post-upgrade-trained model. The synthetic CTs generated from the test set were evaluated using established image quality metrics. Furthermore, dosimetric accuracy was assessed using the patient's clinical treatment plan and our existing clinical NTCP models. RESULTS: Comparison of the average mean absolute error (MAE) between the baseline pre- and post-upgrade (PU) models shows no significant difference. The baseline model exhibited an average MAE of 81.46 ± 49.0 HU and 86.25 ± 14.49 HU for DCNN and cycleGAN, respectively. The TL-05 model demonstrated an average MAE of 69.85 ± 5.9 HU and 95.0 ± 10.95 HU, while the post-upgrade new model had an average MAE of 74.4 ± 12.42 HU and 65.32 ± 10.36 HU for DCNN and cycleGAN, respectively. Additionally, dosimetric quantities showed no significant differences, with mean dose differences ranging from -0.98 ± 3.74% to 2.99 ± 4.74% for DCNN and -0.34 ± 5.45% to 3.15 ± 6.68% for cycleGAN, compared to the post-upgrade new model. Evaluation of the difference between the normal tissue complication probability (∆NTCP) values between the verification CT (rCT) and post-upgrade models showed minimal deviations ranging from -0.001% to -0.03% and 0.0006% to 0.0027% for Grade 2 or higher dysphagia, for DCNN and cycleGAN, respectively. CONCLUSION: Transfer learning strategies, including fine-tuning or freezing feature extraction layers, can minimize disruptions in sCT-dependent workflows. Moreover, the small number of patients required to implement these methods can mitigate extensive downtime due to the limited availability of new data from post-upgrade sources.