Deep learning to segment pelvic bones: large-scale CT datasets and baseline models.

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Pelvic bone segmentation in CT scans is crucial for diagnosing and planning surgeries for pelvic bone diseases.
• Current segmentation methods struggle with variations in CT image appearance, including multi-site domain shifts, contrasted vessels, fractures, and artifacts.
• A lack of large-scale annotated pelvic CT datasets has limited the exploration of deep learning approaches.

Purpose of the Study:

• To address the data gap by curating a large-scale pelvic CT dataset from multiple sources.
• To develop a robust deep learning model for simultaneous multi-class segmentation of pelvic bone structures (lumbar spine, sacrum, left hip, right hip).
• To introduce a novel post-processing technique using the signed distance function (SDF) to enhance segmentation accuracy.

Main Methods:

• Curated a dataset of 1184 pelvic CT volumes with diverse appearance variations.
• Proposed a deep multi-class network trained on images from multiple domains to learn robust feature representations.
• Implemented a post-processor based on the signed distance function (SDF).

Main Results:

• The automatic segmentation method achieved a high average Dice score of 0.987 on metal-free volumes.
• The SDF post-processor significantly reduced the Hausdorff distance by 15.1% compared to traditional methods.
• Experiments demonstrated the effectiveness and robustness of the proposed method across various image challenges.

Conclusions:

• The developed large-scale dataset and deep learning model significantly advance pelvic bone segmentation accuracy.
• The publicly released dataset, code, and models aim to foster further research and development in the medical imaging community.
• This work provides a valuable resource for improving clinical diagnosis and surgical planning related to pelvic bone diseases.