Two-Stage Structure-Focused Contrastive Learning for Automatic Identification and Localization of Complex Pelvic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedic Surgery

Background:

• Pelvic fractures are severe traumas with high morbidity and mortality rates.
• Accurate diagnosis and surgical planning require effective identification and localization of complex pelvic fractures.
• Existing methods struggle with the complexity of pelvic fractures, including multiple fragments and irregular morphology.

Purpose of the Study:

• To develop a novel two-stage method for automatic identification and localization of complex pelvic fractures.
• To overcome limitations of existing methods by incorporating pelvic anatomy symmetry properties.
• To improve accuracy in diagnosing and planning surgeries for pelvic fractures.

Main Methods:

• Developed a two-stage method combining pelvic symmetry properties and feature differences.
• Implemented supervised contrastive learning with a Siamese deep neural network and a structural attention mechanism.
• Introduced a structure-focused attention (SFA) module to capture spatial structural features and enhance fracture zone recognition.

Main Results:

• Achieved a mean accuracy of 0.92 and sensitivity of 0.93 on 103 clinical CT scans from the CTPelvic1K dataset.
• Outperformed three state-of-the-art contrastive learning methods and five advanced classification networks.
• Demonstrated effectiveness in identifying and localizing various types of complex pelvic fractures.

Conclusions:

• The novel two-stage method effectively identifies and localizes complex pelvic fractures using CT scans.
• The approach, utilizing pelvic symmetry and a Siamese deep neural network with SFA, shows superior performance compared to existing methods.
• This technique holds promise for enhancing diagnostic accuracy and surgical planning in pelvic fracture cases.