Design And Implementation Of ABUS Video Fast Segmentation Algorithm Based On Deep Feature Flow

Breast cancer is one of the most threatening diseases to women's health.As a breast cancer detection machine of automatic scanning,ABUS(Automated Breast Ultrasound)reduces the subjectivity of the scanning,but gives the doctor serious burden of reading.One examination of ABUS machine produces 6-10 video,and each video has 3300 images.In order to realize the second level diagnosis of each video,using the computer aided diagnosis technologies to locate the tumors in ABUS videos fleetly and accurately has important reference significance on the doctor's subsequent diagnosis.At present,many computer technologies have been focused on the coronal plane videos of ABUS.However,the axis plane videos contain original information in the ABUS machine,but the axis plane videos are complex,and images of axis plane contain more noises.What's more,the proportion of tumor in the axis plane videos is small,so the tumor segmentation of axis plane is a huge challenge.DFFGA(Deep Feature Flow of Group convolution of Attention mechanism)model is proposed in this paper,and we tested DFFGA model in the ABUS datasets of 100 axis plane videos.In detail,We improved Deep Feature Flow based on the three aspects.(1)The key frame selection network is designed to realize the adaptive key frame selection because of the artificial setting in Deep Feature Flow,and the low level features of each frame are calculated.Based on the difference of low level feature of current frame and key frame to decide the new key frame.The key frame selection can be applied to different data sets.(2)This paper introduced channel and spatial attention mechanism to amplify the importance of meaningful channels and positions.By extracting the global context features the network gives different weights to channels.Similarly,spatial attention module through the funnel structure gives positions of feature map different weights.The implementation of attention mechanism will focus on the important characteristics of the channel and important positions of position,in order to enlarge the importance of meaningful channels and locations to achieve the purpose of accurate segmentation.(3)The whole network is lightened to realize the fast segmentation of axisplane videos,and the deep separable convolution is used to decompose ordinary convolution into deep convolution and point-type convolution,which reduces the floating point computation of the model.As a result,the DFFGA model reached 79.66% of Io U,85.62% of Acc,87.83%of TPR and 12.17% of FPR in ABUS datasets,which not only guaranteed the execution efficiency,but also improved the segmentation accuracy of the model.