Abstract：Abstract: Aiming at the problems that echocardiogram with many noises, weak boundaries, and difficulty in defining the cardiac contours, a Spatial Frequency Divided Attention Network (SFDA-Net) for ultrasound image segmentation was proposed to obtain the four-chamber heart more accurately. With the help of octave convolution, the high and low frequency parallel processing of image was constructed and used in the entire network to get diverse information. At the same time, the Convolutional Block Attention Module (CBAM) was added to allow pay more attention to the target area that needs to be segmented in image feature recovery, so as to reduce the loss of the entire target area. Besides, focal tversky loss was considered as a outstanding loss function to reduce the weight of simple samples and strengthen the attention on difficulty samples, lessen the error introduced by pixel misjudgment between each category, Through multiple sets of comparative experiments, the parameters of the proposed method are lower than the original UNet++, but the segmentation accuracy is increased by 6.2%, dice sore is rised by 8.76%, Mean Pixel Accuracy (mPA) is ascended to 84.09%, Mean Intersection Over Union (mIoU) is increased to 75.6%. SFDA-Net is steadily improve network performance while reducing parameters, making echocardiographic segmentation more precise.
