Bone segmentation on whole-body CT using convolutional neural network with novel data augmentation techniques.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate bone segmentation is crucial for analyzing whole-body CT scans.
• Existing methods may lack robustness across diverse imaging conditions.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) algorithm for automated bone segmentation on whole-body CT.
• To assess the algorithm's performance and generalizability across different datasets.

Main Methods:

• A U-Net based convolutional neural network (CNN) architecture was employed for bone segmentation.
• Three datasets were utilized for training and validation, including in-house, The Cancer Imaging Archive, and public datasets.
• Data augmentation techniques (conventional, mixup, RICAP) were evaluated to enhance model performance and robustness.

Main Results:

• The CNN model achieved high accuracy, with a mean Dice coefficient of 0.983 on the in-house dataset and 0.943 on a secondary dataset.
• The model trained on public data also demonstrated strong performance (mean Dice coefficient of 0.947).
• Data augmentation methods, particularly conventional techniques and RICAP, proved effective in improving segmentation.

Conclusions:

• The developed CNN-based model provides accurate bone segmentation for whole-body CT with generalizability.
• Data augmentation is essential for building robust models, especially when working with smaller datasets.
• This approach offers a reliable tool for bone segmentation in various clinical and research settings.