Multi-scale Target Recognition Of 3D Ultrasound Images Based On Attention

With the development of computer hardware and software infrastructure,as well as the accumulation of an increasingly digital society brought about by the drive massive amounts of data,research on deep learning based on data-driven development has been rapidly developed.At the same time,in the field of medical imaging,with the development of medical imaging technology,increasingly refined medical imaging can provide a large amount of valuable information to help doctors make more accurate diagnosis of the disease.Breast cancer is the cancer with the highest incidence and highest mortality in the female population.Early detection and treatment can effectively reduce the mortality rate of breast cancer,and breast ultrasound examination as an effective inspection method is also widely used in most countries and regions.Automatic breast three-dimensional volumetric ultrasound scanning as a new medical diagnostic technology can automatically scan the entire breast and generate three-dimensional images.The doctor can view the patient's image at any time on the standardized image platform,and effectively improve the detection rate of early breast cancer.However,medical resources are severely scarce and unevenly distributed in many countries and regions.The final diagnosis results are heavily dependent on the doctor's subjective experience,and the complete diagnosis takes a lot of time.Based on the clinical high-quality breast three-dimensional volumetric ultrasound data provided by the hospital,this study constructed a computer-assisted breast lesion recognition model based on the deep learning attention mechanism to provide a key support for further breast cancer computer automatic diagnosis/screening.Compared with traditional two-dimensional image research,this study deals with threedimensional images.According to the characteristics of the problem,this study mainly analyzes from the point of view that a single data sample is large and the target volume is very small.Aiming at the large volume of a single sample of a three-dimensional volumetric image,inspired by the doctor's clinical diagnostic process,the "two-phase model of threedimensional image target detection based on the attention mechanism(3DMS A-CNN)" was proposed as the overall framework.Aiming at the existence of a large number of small lesions in breast cancer,a model of "Multi-Scale Focal Lesion Detection and Localization Based on Three-Dimensional Convolution" was proposed.In order to reduce the false positive rate of the target recognition and improve the sensitivity of the identification of malignant lesions,the ‘Three-dimensional convolutional neural network based lesion subclassification' model was proposed.The multi-scale lesion detection and localization model based on three-dimensional convolution is used as the first stage.The model extracts the multi-scale features of the input image and fuses the multi-level features through interpolation to improve the semantic expression ability of the shallow features.In order to improve the detection ability of lesions at different scales,dense candidate frame detection was used.The Focal Loss function was used to solve the problem of extreme imbalance of positive and negative samples caused by dense candidate boxes.Due to the limitation of computer hardware,in the process of model implementation,the data is down sampled by the pooling layer before being input into the model.To improve the overall generalization of the model,we modify the Batch Norm layer to synchronized multi-GPU Batch Norm,and change the model from single-precision floating-point to half-precision floating-point.A fine-grained classification model based on three-dimensional convolutional neural networks is used as the second stage.Through the suspected candidate regions predicted in the first stage,the corresponding regions are obtained from the original image for further subdivision.By optimizing the implementation of 3D Dense Net,efficient use of video memory is achieved,and the number of training samples for a single batch of models is increased.At the same time,in order to enhance the model's generalization ability,data enhancement was performed using the “mixup” algorithm,which effectively improved the model generalization ability of the second stage.In experiments,ABUS image files with 3012 patients provided by the hospital were used.Among the 2479 patients,one or more lesion markers were included in one image file.2245 patients with lesion markers were randomly selected as training set and 767 patients were used as test set.The deep learning framework Pytorch was used to implement the whole model.After sufficient experiments,the sensitivity of the final model to the lesion reached 91.62%,and the specificity reached 85.20%,which initially met the needs of clinical application.