Projection-domain scatter correction for cone beam computed tomography using a residual convolutional neural network

Area of Science:

• Medical Imaging
• Radiology
• Deep Learning

Background:

• Scatter degrades cone beam computed tomography (CBCT) image quality.
• Conventional scatter correction methods rely on analytical models with assumptions, limiting accuracy.
• Developing advanced methods for accurate scatter removal is crucial for quantitative imaging.

Purpose of the Study:

• To develop an effective scatter correction method for CBCT using a residual convolutional neural network (CNN).
• To evaluate the performance of the CNN-based method against conventional techniques.
• To explore transfer learning for adapting the model to different scanning modes (full-fan vs. half-fan).

Main Methods:

• A U-net based 25-layer CNN was trained using Monte Carlo simulated projection data.
• The model was trained end-to-end with two loss functions, employing data augmentation.
• Performance was compared to the fast adaptive scatter kernel superposition (fASKS) method using digital and anthropomorphic phantoms.
• Transfer learning was used to fine-tune the model for half-fan scans.

Main Results:

• The CNN-based method significantly reduced scatter and improved Hounsfield Unit (HU) accuracy compared to fASKS.
• Root mean squared error was reduced to 0.0862 for CNN-corrected projections versus 0.117 for fASKS.
• Image quality metrics (MAE, MSE, PSNR, SSIM) showed superior performance of the CNN method.
• Transfer learning enabled effective scatter correction in half-fan scans with minimal additional data.

Conclusions:

• The proposed deep learning method offers an effective solution for CBCT scatter correction.
• This technique enhances quantitative imaging accuracy.
• It holds significant value for image-guided radiation therapy.