Deep learning based synthetic CT from cone beam CT generation for abdominal paediatric radiotherapy

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiation Oncology
• Deep Learning for Image Synthesis

Background:

• Adaptive radiotherapy requires high-quality computed tomography (CT) images for accurate dose calculations.
• On-board cone beam CT (CBCT) images often lack the necessary quality for direct use in dose re-optimization.
• Deep learning offers potential for synthesizing CT-quality images from CBCT data.

Purpose of the Study:

• To develop and evaluate a deep learning framework for synthesizing CT-quality images from CBCT.
• To improve the quality and structural consistency of synthetic CT (sCT) images for pediatric abdominal radiotherapy.
• To address challenges of limited data and inter-fraction variability in pediatric imaging.

Main Methods:

• Proposed a novel framework using cycle-consistent Generative Adversarial Networks (cycleGANs) for CBCT-to-CT synthesis.
• Incorporated global residuals learning and modified the cycleGAN loss function for enhanced structural consistency.
• Utilized a smart 2D slice selection and weakly paired data approach to leverage diverse patient datasets for training.

Main Results:

• The proposed method demonstrated improved image similarity metrics (lower Mean Absolute Error) compared to baseline cycleGAN.
• Achieved higher structural agreement for gastrointestinal gas (improved Dice Similarity Coefficient) between source and synthetic images.
• Showed smaller differences in water-equivalent thickness metrics, indicating better preservation of dose calculation properties.

Conclusions:

• Innovations to the cycleGAN framework significantly enhanced the quality and structural consistency of synthetic CT images.
• The developed method shows promise for improving dose calculation accuracy in adaptive radiotherapy workflows, particularly for pediatric patients.
• The approach effectively addresses data scarcity and anatomical variability challenges in medical image synthesis.