Efficient Automatic Detection of Uterine Fibroids Based on the Scalable EfficientDet

Abstract

Uterine fibroids refer to benign tumors formed by uterine smooth muscle tissue hyperplasia, high frequency in women between 30 and 50 years old. By the age of 50 years, 80% of women have one or more uterine fibroids, and about half of these patients are symptomatic and in need of treatment. It's ranking the third highest incidence of all gynecological diseases. Generally, it is a benign tumor, but it can also have certain effects on women's bodies, such as causing infertility. Early detection and treatment are essential measures to reduce morbidity. Ultrasound is the preferred imaging method, and with the continuous development of deep learning in the field of medical image analysis, many applications related to object detection have good performance. Computer-assisted diagnosis can further solve the subjective uncontrollability problem caused by different doctors' reading films. Because doctors' inexperience and fatigue can reduce the diagnostic accuracy of uterine fibroids, this paper proposes a scalable EfficientDet to detect the ultrasound images of uterine fibroids and uses the Convolutional Neural Network (CNN) to extract their features. The backbone network uses EfficientNet, and then it is used together with BiFPN to improve the accuracy of the model. This method can not only benefit non-professional ultrasonologists but also provide sufficient auxiliary diagnostic effects for high-quality ultrasonologists to provide a reliable basis for future treatment and surgical resection. Finally, the effectiveness of this method is experimentally compared with other existing methods. Our method has an average accuracy of 98.88% and an f1-score of 98%. We demonstrate that the methods of this study are superior to other neural networks. And it can bring sufficient benefits to ultrasonologists. We summarize and analyze various detection algorithms, and discuss their possible future research hotspots.