Adnexal torsion diagnosis framework with CT-based adaptive preprocessing and deep neural networks.
Authors
Affiliations (6)
Affiliations (6)
- School of Computer Science and Engineering, Kyungpook National University, Daegu, 41566, Republic of Korea.
- Department of Obstetrics and Gynecology, School of Medicine, Kyungpook National University, Kyungpook National University Chilgok Hospital, Daegu, 41404, Republic of Korea.
- Department of Obstetrics and Gynecology, School of Medicine, Kyungpook National University, Kyungpook National University Hospital, Daegu, 41944, Republic of Korea.
- Department of Radiology, School of Medicine, Kyungpook National University, Kyungpook National University Chilgok Hospital, Daegu, 41404, Republic of Korea.
- Department of Obstetrics and Gynecology, School of Medicine, Kyungpook National University, Kyungpook National University Hospital, Daegu, 41944, Republic of Korea. [email protected].
- School of Computer Science and Engineering, Kyungpook National University, Daegu, 41566, Republic of Korea. [email protected].
Abstract
Adnexal cystic torsion is a gynecological emergency that requires prompt and accurate diagnosis followed by immediate surgical intervention to preserve ovarian function. Although ultrasound or computed tomography (CT) are commonly employed, their diagnostic accuracy is often limited by image quality and interobserver variability. This study aimed to evaluate the feasibility of applying deep learning models for the automated diagnosis of adnexal torsion using abdominopelvic CT. We retrospectively collected 1,191 CT scans from 514 women at two tertiary hospitals in Korea. The disease group included 259 surgically confirmed cases of adnexal torsion, while the control group comprised 255 cases of adnexal cysts without torsion, evaluated during diagnostic exploration for acute abdominal pain. Preprocessing included histogram-based clustering of Hounsfield unit distributions for adaptive windowing and anatomical region selection using the pretrained segmentation model TotalSegmentator. We trained both slice-based 2D multiple instance learning frameworks and fully volumetric 3D convolutional neural networks were using stratified five-fold cross-validation. Among the models tested, the 3D EfficientNet architecture demonstrated the best performance, with 79.25% AUC, 73.4% accuracy, 73.77% specificity. These findings highlight the potential of deep learning-assisted CT interpretation as a tool in gynecologic emergency settings.