Research On Deep Learning-based Segmentation Methods In Ultrasound Images

Because of its benefits of being non-invasive,radiation-free,inexpensive,practical,and real-time imaging,medical ultrasound imaging systems are widely employed in clinical diagnosis and therapy.It takes ultrasound doctors a number of years to complete their formal education and clinical experience accumulation due to the complex and changing anatomical structure of human organs,the proper guidance of ultrasound probes,the interception of standard scanning sections,the marking of tissue areas,and other conventional ultrasound inspection processes.As a result,identifying and comprehending ultrasound images is a challenging problem that is both the clinical technical operating norm for ultrasound specialists and the primary focus of research for intelligent ultrasound imaging systems.Ultrasound image segmentation is the main study topic of ultrasonic image recognition and interpretation and a crucial phase in the process of computer-aided intelligent ultrasonic diagnosis.The segmentation algorithm can help the ultrasound doctor discover areas of interest and reduce their workload by automatically segmenting human organs,tissues,and focal areas in the ultrasound image.The automatic segmentation method for ultrasonic images faces significant obstacles as a result of the inherent noise and artifacts in these images,including inconsistent texture in the target area,similar background texture area,and blurring or missing edge of the target area.To alleviate the above problem,this thesis first constructs the ultrasound image segmentation process as a sequential task with gradual difficulty,and then designs a learning strategy based on the guidance of the pixel-level gradual difficulty curricula to complete the segmentation of the target area from easy to difficult.A data-driven course mining method,which can adaptively customize learning courses for different application scenarios of medical image segmentation,is developed in order to avoid time-consuming and laborious manual labeling of courses and to improve the universality of the image segmentation algorithm.In addition,this thesis also develops a gradual recurrent network for the curricula,which has progressively harder content.In the two ultrasound image segmentation tasks of thyroid and breast lesion segmentation,the effectiveness of gradual recurrent network is confirmed.The proposed method can against noise and artifact interference and increase the accuracy of ultrasound image segmentation when compared to other previous image segmentation methods.Secondly,using the complementary information between the target foreground and background in the ultrasound image,a new image segmentation algorithm is designed,that is,the interactive attention network based on the complementary information,which enhances the feature expression of the foreground and background,strengthens the discrimination ability of the segmentation model,and further improves the representation ability of the algorithm.Three levels information interaction mechanisms are constructed to fully exploit the complementary information between foreground and background features,and the difficulty progressive learning strategy is introduced.The experimental results demonstrate that this method performs better than the gradual recurrent network in the two tasks of thyroid nodule and breast lesion segmentation.Finally,the ultrasound image segmentation method is applied to the fetal head circumference measurement task,and an elliptical guided multitask network is proposed.To increase the edge detection accuracy of fuzzy or missing regions,the network integrates the location information from region segmentation with the detailed information from edge detection.The ellipse constraint is introduced into the model as a priori knowledge of fetal head circumference shape,and force edge detection and region segmentation results approximate the ellipse.To further enhance the accuracy of the fetal head circumference measurement,some modules in the above segmentation approaches are integrated with the fetal head circumference measuring method.A public data set for measuring fetal head circumference is used to confirm the validity of the proposed method.In conclusion,two ultrasound image segmentation algorithms based on deep learning are achieved by combining the from-easy-to-hard strategy of curriculum learning in this thesis.Experimental results demonstrate the proposed methods perform better than the state-of-the-art approaches and are also used to complete the task of measuring the fetal head circumference,which illustrates the potential uses for the segmentation algorithm.