Automatic Diagnosis Method For Benign And Malignant Lesion Areas In Ultrasound Images Based On Heterogeneous Multi-branch Network

Breast cancer is one of the most important cancers threatening women’s health worldwide.ultrasound(US)is one of a primary imageological examination and preoperative assessment for breast nodules.However,in the field of ultrasound diagnosis,it relies heavily on the experience of physicians due to the overlapping image expression of benign and malignant breast nodules.The diagnostic accuracy of physicians with different qualifications differs by up to 30%.Therefore,it is easy to lead to misdiagnosis and increase the needless rate of puncture biopsy.In recent years,with the continuous progress of computer science and the continuous enrichment of medical data sets,computer-aided medical diagnosis has gradually become a hot spot of current research.While the current computer-assisted breast ultrasound diagnosis requires lots of human interactions while the accuracy is not reliable enough.In this paper,an end-to-end model is proposed for automatically nodule classification.We presents a heterogeneous three-branch network(HTBN)for benign and malignant classification of the breast ultrasound images.In HTBN,the image information including ultrasound images,contrast-enhanced ultrasound(CEUS)images and non-image information including patient’s age and other six pathological features are used simultaneously.On the other hand,we proposes a fusion loss function suitable for this heterogeneous multi-branch network.This loss function uses the minimum hyperspherical energy(MHE)based on additive angular margin loss to improve the classification effect.Among them,additive angular margin loss is mainly used to reduce the distance between classes,and MHE is mainly used to increase the distance between classes.The experimental results show that our fusion loss can further increase the angular margin of learned features and reduce the intra-class distance at the same time.In order to validate our method,a breast ultrasound data set with 1303 cases is collected.On this data set,the average diagnosis accuracy of physicians with five-year qualifications is85.3%.However,the classification accuracy of our method is 92.41%.Through experiments,we confirmed our point of view that by incorporating medical knowledge into the optimization process,adding contrast-enhanced ultrasound images and non-image information to the network,the accuracy and robustness of breast diagnosis are greatly improved.