Efficient cascaded V-net optimization for lower extremity CT segmentation validated using bone morphology assessment.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedic Surgery

Background:

• Semantic segmentation of lower extremity bones from CT scans is crucial for orthopedic visualization, diagnosis, and surgical planning.
• Current deep learning methods face challenges with large field-of-view CT scans, leading to slow performance and high GPU memory demands.

Purpose of the Study:

• To investigate efficient methods for anatomical context representation in deep learning for faster and more accurate bone segmentation.
• To compare the performance of a novel cascaded deep learning approach against state-of-the-art methods.

Main Methods:

• A cascaded deep learning approach using V-net blocks was optimized with varying resolution levels, receptive fields, patch sizes, and input voxel dimensions (128^3-64^3-32^3).
• Six lower extremity bones from two patient datasets were manually segmented for training and validation.

Main Results:

• The optimized cascaded V-net achieved an average Dice coefficient of 0.98 ± 0.01, mean surface distance of 0.26 ± 0.12 mm, and 95th percentile Hausdorff distance of 0.65 ± 0.28 mm.
• This method demonstrated significant improvements over nnU-net, requiring 1/12th training time, 1/3rd inference time, and 1/4th GPU memory.
• Morphometric measurements showed high correlation (ICC > 0.8) with manual segmentations, indicating clinical utility.

Conclusions:

• The proposed multi-stage, cascaded V-net approach offers a highly accurate and efficient solution for lower extremity bone segmentation from CT scans.
• This method significantly outperforms existing state-of-the-art techniques in terms of speed and resource utilization.
• The segmentation quality is sufficient for various clinical applications, enabling accelerated orthopedic workflows.