Multi-purposed diagnostic system for ovarian endometrioma using CNN and transformer networks in ultrasound

Abstract

ObjectiveOvarian endometrioma (OMA) is one of the most common ovarian cysts worldwide, seriously impairing the reproductive function of females. Accurate diagnosis is of great significance for appropriate clinical treatment. Nowadays, the definitive diagnosis of OMA is based on its clinical manifestations, while ultrasound is widely employed as a routine diagnostic modality for OMA. However, the ultrasound diagnosis of OMA, which has various challenges, is contingent upon the expertise and experience of doctors. MethodsTo overcome this, we propose an automated method based on deep learning, which performs cyst segmentation and binary OMA (i.e. OMA and non-OMA) classification simultaneously. The features provided by the segmentation branch are fused with the classification features with the assistance of the attention mechanism. In this manner, the classification branch can better focus on the cyst regions of the image and learn more specific information, thereby improving the accuracy of classification. ResultsWe evaluate the method on an extensive dataset containing 1501 images. The proposed model achieved a classification accuracy of 91.36 % and a Dice score of 85.42 %, which outperformed several popular networks involved in this study. What’s more, further reader study revealed that our model demonstrates comparable performance with that of the senior doctors, while superior to that of the junior doctors. Conclusion and significanceTo the best of our knowledge, this is the first deep learning method specially designed for OMA diagnosis. This indicates that our model has the potential to assist junior doctors in improving their diagnostic accuracy.