Building a patient-specific model using transfer learning for four-dimensional cone beam computed tomography...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy

Background:

• Deep learning models for 4D-CBCT quality augmentation were previously developed using group data.
• Group-based models lacked patient-specific optimization, limiting depiction of small anatomical structures like lung vessels.

Purpose of the Study:

• To improve deep learning model performance for individual patients using transfer learning.
• To enhance the quality of augmented undersampled 4D-CBCT images for clinical applications.

Main Methods:

• A U-Net-based model was initially trained on group data for 4D-CBCT augmentation.
• Transfer learning (layer-freezing and whole-network fine-tuning) was applied to fine-tune the model for individual patients.
• Model performance was evaluated quantitatively (SSIM, PSNR) and qualitatively against ground truth images.

Main Results:

• Patient-specific models qualitatively revealed more detailed lung structures compared to group-based models.
• Quantitative metrics showed significant improvements: SSIM increased from 0.924 to 0.958, and PSNR from 33.77 to 38.42.
• Layer-freezing transfer learning was more efficient, with training times as short as 10 minutes, and augmentation effectiveness increased with fewer projections.

Conclusions:

• Patient-specific models optimized by transfer learning are efficient and effective for improving 4D-CBCT image quality.
• This approach significantly enhances augmented undersampled 3D and 4D-CBCT images.
• The optimized models hold substantial value for image-guided radiation therapy applications.