Deep Learning-Assisted Ultrasonic Diagnosis and Localization of Testicular Appendix Torsion

Condition(s) studied

Study summary

Ultrasound data were both retrospectively and prospectively collected from the primary center and six other sub-centers. Combined with clinical diagnostic outcomes, the data labeling was completed by physicians with extensive clinical experience. In this study, ConvNeXtV2 was used as the classification network and YOLOv12 was adopted as the detection network.The retrospective dataset from the primary center was split into training, validation, and test subsets, on which the model was trained, validated, and tested respectively; additional validation was conducted on both retrospective and prospective datasets from the primary center and sub-centers.Meanwhile, four physicians were assigned to interpret the ultrasound data from the retrospective and prospective datasets from the primary center and sub-centers using two diagnostic methods-independent diagnosis and artificial intelligence (AI)-assisted diagnosis-and the diagnostic accuracy of these two approaches was further compared.By collecting and learning the treatment methods of patients in the primary center training set, predicting the treatment methods of patients in the sub-center datasets, and comparing the proportion of surgeries predicted by AI with the actual proportion of surgeries, the efficacy of the model was verified.

Eligibility

Primary outcome measure(s)

• Accuracy of deep-learning model verify four conditions:testicular appendage torsion;testicular torsion;epididymitis and normal condition — From image input to result generation is expected to be 24 hours
  accuracy of deep-learning model verify four conditions:testicular appendage torsion;testicular torsion;epididymitis and normal condition

Trial sites (1)