Segmentation And Classification Of Breast Ultrasound Image Based On Attention Mechanism And Difficult Sample Mining

Ultrasound diagnosis is often used as the first choice for imaging examination and preoperative evaluation of breast cancer,but there are some problems such as poor quality of data imaging,overlapping of benign and malignant nodule images,heavy dependence on doctors’ experience and more human-computer interaction.Therefore,in order to reduce misdiagnosis and unnecessary biopsy rate and improve the level of diagnosis automation.An end-to-end model based on attention mechanism and difficult sample mining is proposed in this thesis to realize the automatic segmentation of breast nodule area and the identification of benign and malignant.First,after the contrast analysis of speckle noise and low contrast of ultrasonic image,the edge enhanced anisotropic diffusion denoting model is used After that,the focus mechanism is used to segment the nodal region to eliminate the interference of the background to the subsequent classification operation.Finally,an improved loss function is proposed to enhance the discriminant ability for the benign and malignant features of the nodule,mining difficult samples which are easy to be misjudged in dataset by combining shape descriptors,in order to distinguish the difficult samples better,this thesis applies the improved loss function,and on this basis,constructs the shape constraint loss term of hard samples,which is used to adjust the feature mapping between samples with similar shape but different categories,and expand the decision distance between different categories.In order to verify the effectiveness of the above algorithm,this thesis constructs a breast ultrasound data set with 1805 images,on which the average accuracy of doctors with five years’ experience is 85.3%,while the classification accuracy of the method in this data set is92.58%,sensitivity is 90.44%,specificity is 93.72%,AUC is 0.946,which are superior to the comparison algorithm;compared with the traditional softmax loss function The evaluation indexes increased by 5% ～ 12%.The experimental results show that the end-to-end breast ultrasound image segmentation and classification method proposed in this thesis has strong practicability;through the integration of medical knowledge into the optimization model,the increase of difficult sample shape constraint loss items can improve the accuracy and robustness of breast benign and malignant diagnosis,and each evaluation index is higher than that of ultrasound doctors,which has a certain clinical application research value.