Research On Breast Cancer Detection Based On Automatic Breast Ultrasound Images

In recent years,with the increase of breast cancer patients,a convenient,safe,and inexpensive 3D automated volumetric breast ultrasound system(ABUS)has been widely used for breast cancer.ABUS can provide a full range of vision,reduce operator dependence,and greatly help doctors diagnose breast cancer.But ABUS also raises many issues.First of all,clinically,the amount of image data generated is large,meanwhile there are fewer experienced doctors,which leads to low reading efficiency.At present,an automatic diagnosis method is urgently needed to save the time of the entire diagnosis workflow.In the area of 3D ultrasound image research.The breast cancer detection based on ABUS image has many problems such as huge computing resource consumption,poor image quality,and data imbalance.In order to overcome these challenges,this paper proposes a new end-to-end 3D convolutional neural network(CNN)for breast cancer detectionResearching related methods in this field,traditional machine learning methods are difficult to cope with complex imaging conditions,and have low speed and performance cross Meanwhile,the popular deep learning methods have problems such as computational redundancy and slow inference time.Analyzing the problems encountered in the research,this paper proposes the following four methods.First,according to the characteristics of three-dimensional ultrasound imaging,when annotating ABUS volumes,the inverse digital signal converter is used for the gold standard.Thus,we change the input and output of the neural network,reduce the amount of calculation,and make the network easier to learn the signal attenuation.Second,in order to detect breast cancer automatically,we use 3D Inception Unet as the segmentation network and take advantage of the prior knowledge of breast cancer to remove false positive candidates.Third,for the problem that 3D neural networks take plenty memory of GPU and are difficult to train,we design an Inception Block that combines 2D and 3D CNNs to extract features of points,lines,areas,and volumes inspired by the convolution kernel decomposition.Third,in order to reduce the rate of missed diagnosis,this paper proposes an asymmetric segmentation loss function that is oriented to clinical needs,and automatically focuses on difficult to segment small lesions to achieve a balance between specificity and sensitivity.Fourth,according to the characteristics of ultrasound imaging,we use a method based on 3D random walk to calculate a Confidence Map,evaluate the quality of ultrasound imaging,and combine the post-processing method of Confidence Map to further reduce the number of false positivesBenefited from the above improvements,our method achieved good results on ABUS volumes on our 199 positive patient data set,reaching a detection rate of 95.51%and averaging 4.2 false positive candidates per volume(FPs),1.3 FPs on 192 negative patient data.On our end-to-end network,the average inference time of a Volume is 0.1 seconds,which can help doctors quickly locate the tumors.Experiments indicate that the asymmetric segmentation loss function and Inception Unet network make our model more suitable for breast cancer detection tasks of ABUS than other 3D Unet-like neural networks.