Research On Ultrasound Image Segmentation Methods Based On Deep Learning

Ultrasound examination is one of the most widely used medical imaging techniques,with applicability and no ionizing radiation.However,in recent years,the development of automatic segmentation of ultrasound images lags behind that of CT images and MRI images significantly.The reason for this is that the ultrasound artifacts and edge blurring issues bring a limitation to the accuracy of segmentation.Although deep learning-based medical image segmentation methods have achieved excellent performance in many tasks and gradually occupied the leading position in the industry,the development of deep learning segmentation methods in ultrasound is still inadequate,being unable to deal with the mentioned problems.To explore the improvement of deep learning-based ultrasound image segmentation,this thesis presents discussion and analysis from the following three aspects:1.Introducing structured prior information in the segmentation taskThe current segmentation of medical ultrasound images only relies on the information contained in images,which has obvious performance deficiencies because it ignores the use of prior information.In practical application scenarios,ultrasound doctors usually use rich medical prior information to assist in the identification and diagnosis of organs or lesions.Based on this,the thesis proposes to introduce structured medical prior information in the segmentation task to supplement image information to improve the segmentation performance.Based on the decomposition model,the thesis studies the prior information introduction methods based on multi-task learning and the prior information introduction methods based on multi-modal fusion respectively,and validates the methods’ effectiveness through experiments.2.Introducing edge supervision to refine segmentation edgesEdge blurring is one of the most difficult challenges in ultrasound image segmentation tasks.In this thesis,edge degradation experiments are conducted to verify the fatigue of commonly used segmentation frameworks for edge segmentation.Concentrating on this problem,the thesis explores edge enhancement methods based on edge supervision,and two different edge enhancement schemes are applied: edge enhancement methods based on high-frequency decoupling and edge enhancement methods based on a boundary distance metric.The methods are capable to improve the identification of bad edges and restoration of details in ultrasound images.3.Use morphological constraints to compensate for bad edgesConfronted with the lack of attention to the overall morphological features of organs in common deep learning methods,two kinds of solutions are introduced: shape constraint methods based on Fourier descriptors on the one hand,and uses low-frequency Fourier descriptors to introduce the network’s attention to the overall shape features? The use of loss function constrained by the active contour energy where a joint constraint on the target contour and the regional scale is introduced to achieve effective restoration of bad edges.