Deep learning-based segmentation in prostate radiation therapy using Monte Carlo simulated cone-beam computed...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Organ segmentation in cone-beam CT (CBCT) is crucial for adaptive radiotherapy but is challenging due to low image contrast and high inter-observer variability.
• Existing methods like deformable image registration and deep learning require either sensitivity to large deformations or extensive manual segmentation of CBCT images.

Purpose of the Study:

• To develop an alternative approach for training deep learning segmentation models using simulated CBCT images.
• To overcome the difficulties associated with obtaining accurate organ contours from real CBCT images.

Main Methods:

• Generated pseudo-CBCT (pCBCT) images from segmented planning CT images using GATE Monte Carlo simulation.
• Trained nnU-Net models using pCBCT data, real CBCT data, segmented real CT data, and combinations thereof.
• Evaluated segmentation performance using Dice Similarity Coefficient (DSC) and Hausdorff distance on various CBCT and pseudo-CT datasets.

Main Results:

• Models trained with pCBCT achieved segmentation performance comparable to models trained with real CBCT images.
• pCBCT-trained models demonstrated high clinical acceptability, with a majority of segmentations requiring minor or no corrections.
• The pCBCT approach outperformed other methods in most evaluations, showing its robustness and efficiency.

Conclusions:

• Simulated CBCT images can be effectively used to train deep learning segmentation models (nnU-Net).
• This method circumvents the need for time-consuming and complex manual delineations on actual CBCT images.
• The proposed approach offers a viable solution for improving organ segmentation in radiotherapy planning.