부산대학교 도서관

Infertility, the failure to conceive within a year despite regular sexual intercourse, affects tens of millions of people around the world. Hysterosalpingography (HSG) is a routine imaging tool, can effectively evaluate the patency of the fallopian tubes. This study aims to explore the potential of artificial intelligence (AI) to automatically interpret images using convolutional neural network (CNN) and transfer learning technology. The imaging data of infertility patients who underwent HSG examination in our hospital were retrospectively included, and were interpreted and annotated by two radiologists. The performance of the model is optimized through transfer learning, data amplification technology, and various hyperparameter settings to perform binary classification predictions of both patency or any one side obstruction. The research results show that using 730 limited training data, the model trained with shufflenet v2 achieved an accuracy of 90%, a recall rate of 93%, an F1 value of 93%, and a precision of 94% in the 91 test dataset. The specificity was 82% and the negative predictive value was 78%. Through the analysis of heat map, the model demonstrated the ability to correctly identify and focus on the position and shape of the fallopian tube in the image. In summary, the CNN model is feasible and effective in automatically classifying whether the fallopian tube is patent in HSG images. Through further research and optimization, we hope that the model can be actually used in clinical practice in the future to practice artificial intelligence-assisted medical treatment.