Two-stage multi-task deep learning framework for simultaneous pelvic bone segmentation and landmark detection from CT...

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Orthopedic Surgery

Background:

• Accurate pelvic bone segmentation and landmark identification are vital for preoperative planning in total hip arthroplasty.
• Diseased pelvic anatomy often compromises the precision of standard segmentation and detection methods, potentially leading to surgical complications.

Purpose of the Study:

• To develop and validate a novel two-stage, multi-task algorithm for improved pelvic bone segmentation and landmark detection.
• To enhance accuracy specifically in cases with diseased pelvic anatomy, addressing a critical limitation in current clinical practice.

Main Methods:

• A coarse-to-fine, two-stage framework employing multi-task learning for simultaneous segmentation and landmark detection.
• The first stage uses a dual-task network for global analysis, followed by an edge-enhanced dual-task network in the second stage for local refinement and boundary delineation.

Main Results:

• The algorithm achieved high Dice Similarity Coefficient (DSC) scores for pelvic structures (e.g., 0.97 for left/right hips) and an average landmark error of 3.24 mm.
• The second stage significantly improved acetabular boundary segmentation accuracy by 5.42% over existing methods, outperforming state-of-the-art approaches.
• The entire segmentation and detection workflow was completed in approximately 10 seconds, demonstrating computational efficiency.

Conclusions:

• The proposed multi-task, coarse-to-fine strategy significantly enhances pelvic bone segmentation and landmark detection accuracy, particularly for diseased hip images.
• This advancement facilitates more precise and rapid preoperative planning for acetabular cup prosthesis design and implantation.