Task-based transferable deep-learning scatter correction in cone beam computed tomography: a simulation study.

Area of Science:

• Medical Physics
• Radiology
• Image Processing

Background:

• X-ray scatter degrades image quality in cone beam computed tomography (CBCT).
• Convolutional neural networks (CNNs) show potential for scatter correction but face challenges with large datasets and generalizability.
• Task-based paradigms offer a solution to enhance CNN application in scatter correction.

Purpose of the Study:

• To introduce a task-based paradigm for CNN-based X-ray scatter correction in CBCT.
• To overcome limitations of extensive datasets and model generalizability in CNN scatter correction.
• To enhance the application of CNNs for improved CBCT image quality.

Main Methods:

• A U-net architecture CNN was employed with a two-stage training process.
• CNNs were pre-trained on geometric phantom projections and fine-tuned using transfer learning (TL) on anthropomorphic projections.
• 2D scatter ratio (SR) maps were used as targets for scatter prediction and CNN retraining for specific tasks.

Main Results:

• Pre-training achieved accurate scatter ratio (SR) predictions.
• Transfer learning (TL) further improved SR prediction accuracy with significantly less data and faster retraining times.
• CNN models demonstrated successful X-ray scatter correction in anthropomorphic structures.

Conclusions:

• A fast, low-cost methodology for task-specific CNN development in CBCT scatter correction was established.
• The proposed methodology enables effective scatter correction in CBCT using minimal data.
• The developed CNN models show promise for correcting scatter in previously unseen anatomical structures.