Surgical planning of pelvic tumor using multi-view CNN with relation-context representation learning

Area of Science:

• Musculoskeletal Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Limb salvage surgery for malignant pelvic tumors is complex due to anatomical challenges.
• Accurate tumor resection with margins is critical but hindered by inefficient image planning.
• Current methods lack efficient and repeatable tumor identification and segmentation in medical imaging.

Purpose of the Study:

• To develop a novel deep learning-based method for accurate segmentation of pelvic bone tumors in MRI.
• To improve the efficiency and accuracy of surgical planning for limb salvage surgery.
• To address the lack of efficient image planning methods in hospitals.

Main Methods:

• A multi-view fusion network was employed to extract pseudo-3D information from MRI scans.
• Relational context learning was used to enhance feature representation and utilize spatial information.
• The method was evaluated on two independent datasets (90 and 15 patients).

Main Results:

• The deep learning method achieved segmentation accuracy comparable to expert annotation.
• It outperformed several existing segmentation methods on independent datasets.
• Average segmentation time decreased from 1820.3 seconds to 19.2 seconds (approx. 100x reduction).

Conclusions:

• The proposed deep learning method offers a significant advancement in pelvic tumor segmentation accuracy and speed.
• Integrating this method into surgical planning workflows dramatically reduces planning time (15 minutes vs. 2 days).
• This approach holds potential to improve outcomes in limb salvage surgery for malignant pelvic tumors.