Development of an image-based deep learning algorithm to predict ovarian torsion.

BACKGROUND: The aim of this study was to assess the feasibility of developing a deep learning-based algorithm for diagnosing adnexal torsion that incorporates clinical features, sonographic findings, and sonographic images. METHODS: For the development of the machine learning algorithms, retrospective data from 2013 to 2022 were obtained for patients who underwent operative management for concern for adnexal torsion within a single healthcare system. These data were divided into training, validation, and test cohorts. Three machine learning algorithms were developed: (1) decision tree model, which used clinical and sonographic findings; (2) simple neural network algorithm, which used clinical and sonographic findings; and (3) image neural network algorithm, which used clinical features, sonographic findings, and sonographic images. The primary outcome was performance of the models in predicting adnexal torsion, as measured by area under the receiver operating characteristic curve (AUC). RESULTS: 477 patients were identified as possible candidates. 222 patients met inclusion criteria. Of these, 157 patients had adnexal torsion, and 65 did not have adnexal torsion, based on operative findings. The decision tree model had the test AUC 0.5829. The simple neural network model had the test AUC of 0.7572. The image neural network model had the test AUC of 0.8053. CONCLUSION: This feasibility study demonstrated that the image neural network algorithm had improved detection of adnexal torsion compared to the decision tree model or simple neural network model. Further development and refinement of this neural network algorithm could assist in the clinical decision-making regarding adnexal torsion.