Deep-learning-based multiple organ segmentation for CT scout images: applications to automatic CT planning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography (CT) scout images are crucial for planning scan ranges and optimizing radiation dose.
• Manual CT planning is inconsistent and can lead to excessive radiation exposure.
• Accurate scan range prediction is vital for efficient and safe CT examinations.

Purpose of the Study:

• To develop and evaluate a multi-organ segmentation system for CT scout images.
• To enable automatic prediction of scan ranges for CT planning.
• To reduce variability and radiation dose in CT procedures.

Main Methods:

• Trained U-Net models with ResNet-50 encoders to segment 35 anatomic structures from frontal and lateral scout images.
• Utilized organ boundaries from segmentation to automatically determine scan ranges for chest-abdomen-pelvis and abdomen-pelvis CT protocols.
• Evaluated model performance using Dice scores and assessed the impact on scan range prediction accuracy and length.

Main Results:

• Achieved high mean Dice scores for organ segmentation (0.823±0.014 frontal, 0.784±0.013 lateral).
• Predicted scan ranges covered ground-truth anatomy in 87%-100% of cases.
• Reduced scan limit variability and scan lengths in multiple prediction tasks.

Conclusions:

• The developed models accurately segment diverse organs from CT scout images.
• Demonstrated significant potential for improving automatic CT scan range prediction.
• Offers a promising approach for enhancing automated CT planning and radiation dose optimization.