Research On Ultrasound Image Segmentation Algorithm Based On Improved U-Net

Image segmentation is a research hotspot in the field of medical image processing.Fast and automatic segmentation of medical images helps to improve the efficiency of doctors and plays an important supporting role for disease diagnosis.Among the common types of medical imaging,ultrasound has become the first choice of patients and doctors for intraoperative diagnosis because of its non-invasive,real-time,non-radioactive damage and low cost features,but the phenomenon of many noisy,unclear edges and inconsistent target shapes in ultrasound images is common,and most traditional image segmentation methods require manual intervention and are very sensitive to noise,and the segmentation results obtained are not sufficient to meet the demand.With the continuous development of deep learning,the image segmentation task by feature extraction has become mainstream,and the problem of noisy ultrasound images can be well solved,but most of the mainstream network models focus on the improvement of accuracy,thus ignoring the significant increase of computational complexity,which makes the model cannot be easily embedded into modern medical devices.Therefore,in order to reduce the number of model parameters and computational effort while improving the segmentation accuracy of the model,as well as to better cope with the problem of multi-scale variation of segmentation targets in ultrasound images and reduce the sensitivity of the model to noise,this paper improves on the classical encoder-decoder structure,and the main work can be summarized as follows:(1)A feature fusion network based on attention and null convolution is proposed for the problem of noisy and multi-scale variation of segmentation targets in ultrasound images.The network uses U-Net as the base model,and firstly introduces the Efficient Channel Attention Network as the attention mechanism in the encoder part to increase the attention of the model to the segmentation target and reduce the influence of noise information on the segmentation accuracy,and then introduces the space pyramidal pooling of voids module in the jump connection to expand the feature extraction process through the convolution of voids with different expansion coefficients.Finally,a comparison experiment is conducted with other segmentation networks on two ultrasound image datasets to demonstrate that the proposed model has higher segmentation accuracy and is more accurate for target morphology segmentation.(2)A depth-separable network based on a lightweight attention mechanism is proposed to address the problem that the excessive computational complexity of the U-Net++network prevents its easy migration to medical devices.First,considering that the excessive number of parameters and computation in the network is caused by convolutional operations,a diversified depth-separable convolutional layer is proposed instead of the convolutional module in order to maintain a balance between computational complexity and segmentation accuracy.Second,a lightweight coordinate attention mechanism is introduced in the jump connection,which can effectively increase the model’s attention to the segmentation target location information by decomposing the channel attention into two dimensions.In addition,in order to maximize the utilization of the feature information extracted by the network model,a bidirectional jump connection structure is proposed,and a reverse long connection connecting the output of the decoding side and the input of the encoding side is added.Finally,the idea of feature fusion module in U~2-Net is borrowed,and the main structure of U-Net++is improved according to the characteristics of segmentation networks.Compared with U-Net and other classical improved networks on two publicly available datasets,the model proposed in this paper both reduces the computational complexity and improves the segmentation accuracy,and has the optimal segmentation effect.(3)A multiscale fusion network based on residuals and multidimensional convolutional attention is proposed to address the problem of UNe Xt network with low segmentation accuracy in simple datasets.The model improves the convolution phase of the UNe Xt network.Firstly,a multidimensional convolutional attention module is proposed to be added in the jump connection of the model to filter out unimportant information such as noise by fusing the attention weights of different dimensions.Secondly,in the decoder part,a residual-based multiscale fusion module is proposed to replace the convolution module for the problem of variable target morphology in ultrasound images,using parallel null convolution and convolution kernels of different sizes to obtain feature information under different perceptual fields.Finally,a new loss function is used to help the model better segment the detail information in the image.The experimental results show that the proposed model significantly improves the segmentation accuracy with a small increase in computational complexity,and has better segmentation effect on the detail information in ultrasound images.