Adnexal torsion diagnosis framework with CT-based adaptive preprocessing and deep neural networks

Area of Science:

• Radiology
• Artificial Intelligence
• Gynecologic Imaging

Background:

• Adnexal torsion is a critical gynecologic emergency requiring swift diagnosis and surgery to preserve ovarian function.
• Current diagnostic methods like ultrasound and CT have limitations due to image quality and variability.
• Automated diagnostic tools are needed to improve accuracy and efficiency in emergency settings.

Purpose of the Study:

• To assess the feasibility of deep learning models for automated adnexal torsion diagnosis.
• To evaluate the performance of AI models using abdominopelvic CT scans.
• To explore AI's potential in supporting clinical decisions for gynecologic emergencies.

Main Methods:

• Retrospective analysis of 1,191 abdominopelvic CT scans from 514 women.
• Data included surgically confirmed adnexal torsion and control cases (adnexal cysts without torsion).
• Preprocessing involved adaptive windowing and anatomical segmentation; models included 2D multiple instance learning and 3D convolutional neural networks (3D CNNs).

Main Results:

• The 3D EfficientNet architecture, a type of 3D CNN, achieved the highest performance.
• The best model demonstrated an Area Under the Curve (AUC) of 79.25%, accuracy of 73.4%, and specificity of 73.77%.
• These results indicate strong potential for AI in adnexal torsion detection.

Conclusions:

• Deep learning models can be effectively applied for automated adnexal torsion diagnosis using CT.
• AI-assisted CT interpretation offers a promising tool for gynecologic emergency management.
• Further development could enhance diagnostic accuracy and patient outcomes in acute gynecologic conditions.