Room: AAPM ePoster Library
Purpose: To further improve the performance of a deep-learning based CBCT augmentation technique by using transfer learning and patient-specific information.
Methods: A Unet model was developed to augment the quality of undersampled CBCT images from a 3D/4D CBCT scan. The model was first trained using a group of patients’ data, and then was fine-tuned using transfer learning and the patient-specific data. Two transfer learning methods were explored, including layer freezing and whole-network fine-tuning. The augmented CBCT with transfer learning was evaluated by comparing against the groundtruth images both qualitatively and quantitatively using structure similarity index matrix (SSIM) and peak signal-to-noise ratio (PSNR). The performance of patient-specific transfer learning model was compared against the original group based UNet model. The effect of projection number on the model performance was also studied.
Results: Qualitatively, transfer learning effectively improved the fidelity of detailed anatomical structures in the augmented CBCT, compared to the UNet model alone. Quantitatively, transfer learning improved SSIM from 0.924 to 0.958, PSNR from 33.77 to 38.42 for whole volumetric images. The two transfer learning methods had comparable performance in augmenting the images while the layer-freezing method was more time-efficient with training time as short as 10 minutes. The enhancement by transfer learning became more prominent as the number of projections used for CBCT reconstruction decreased to as low as 90.
Conclusion: The transfer learning method is efficient and effective in enhancing the performance of deep-learning methods for augmenting 3D/4D-CBCT images, which are valuable in IGRT applications. The study also demonstrated the potential of using transfer learning to retrain a group based deep-learning model into a patient-specific model to enhance its performance for the individual patients in various clinical tasks.
Funding Support, Disclosures, and Conflict of Interest: Supported by NIH grants R01-CA184173 and R01-EB028324.