Quality Assessment Method Of Fetal Ultrasound Plane Based On Multi-task

Ultrasound examination is safe and radiation-free and is currently recognized as preferred imaging approach for prenatal diagnosis and birth defect screening.An important step in prenatal ultrasound screening is to evaluate the quality of the plane images collected by sonographers.This step is an important means to ensure the accuracy and standardization of prenatal ultrasound screening and can play a decisive role in the accuracy of biological parameter measurement.Clinically,quality assessment is to identify key anatomical structures in fetal ultrasound plane images based on their own experience and professional knowledge by professional sonographers to determine whether the plane is a standard plane.However,this method will lead to quality assessment that relies heavily on the experience of sonographers,is difficult and has a low degree of standardization,and is inefficient in obtaining planes,resulting in a high rate of misdiagnosis and missed diagnosis in fetal malformation screening.With the gradual maturity and wide application of deep learning technology,the research in the field of prenatal ultrasound medical treatment based on deep learning has become the current and future research trend,but it also brings many challenges.This work explores deep learning-based object detection and semantic segmentation algorithms,and introduces its principles,and then proposes an end-to-end intelligent assistance solution for the task of quality assessment of fetal ultrasound planes.The main contents are summarized as follows:(1)In order to accurately assess the quality of fetal ultrasound plane images,this paper designs and studies a simple but efficient multi-task-based fetal ultrasound plane quality assessment model:MTQANet.The network includes a shared feature extraction network,a feature fusion network,and three branch networks used to solve specific tasks.The three branch networks are respectively used to complete the task of classification of planes,detection of key anatomical structures and segmentation of key structures to be measured in the plane.(2)In view of the fact that there are many key anatomical structures with small scales in fetal ultrasound planes,this paper designs a multi-scale prediction solution,adding an additional feature map that downsamples the input image by 4 times to the feature fusion network of MTQANet for feature fusion,and the original 3-scale detection layer is extended to 4-scale detection layer.The added feature map has a larger scale and a smaller receptive field than the existing feature map,which enables the network to obtain more accurate location information of small-scale targets,thereby effectively improving the detection accuracy of key anatomical structures at smaller scales.(3)In order to effectively utilize the location information of each key anatomical structure in the plane,this paper adds a location attention mechanism that integrates location information into channel attention in the feature fusion mechanism of MTQANet.This mechanism enables the model to focus on the positional relationship of each key anatomical structure,which can improve the detection effect of key anatomical structures.(4)This paper conducts large-scale experiments on 18,000 datasets including upper abdominal transverse sections,thalamus horizontal transverse sections,and femoral long-axis sections to verify the effectiveness of the method proposed in this paper and provides the possibility for future clinical promotion and use.