Research On Breast Ultrasound Image Segmentation Methods Based On Deep Learning

Breast cancer is ranked as the second-highest in diagnosis and fifth in mortality.Early detection and diagnosis play an important role in the cure of breast cancer.Ultrasound image has been the common tool for breast cancer diagnosis because it is inexpensive and radiationless.Ultrasound image segmentation can delineate the tumors automatically.It is the foundation of breast tumor quantification,which is crucial for breast cancer diagnosis.However,traditional methods experience three challenges:(1)Traditional level set methods cannot fully understand the tumor regions with complex characteristics by only low-level features,(2)traditional methods ignore the trade-off between segmentation accuracy and efficiency,(3)It’s difficult for a single model to learn enough useful information of complex tumor regions,result in limitation of segmentation performance.In recent years,deep learning methods have achieved remarkable breakthroughs in medical image segmentation tasks due to their powerful feature representation capabilities.In order to overcome the limitations of existing methods and achieve accurate breast tumor segmentation,combined with deep learning technology,this thesis studies the key technologies of breast ultrasound image segmentation:1.Traditional level set methods cannot fully understand the tumor regions with complex characteristics by only low-level features.In order to address this issue,this paper proposed a contextual level set method for breast tumor segmentation.Firstly,an encoder-decoder architecture network is developed to learn high-level contextual features with semantic information.After that,the contextual level set method has been proposed to incorporate the novel contextual energy term.The proposed term has the ability to embed the high-level contextual knowledge into the level set framework.The learned contextual features with semantic information can provide more discriminative information,which has been directly associated with category labels.Therefore,it is robust to serious intensity inhomogeneity,which is helpful to improve segmentation performance.The experiments took place with the help of three databases,which indicates that the proposed method outperformed traditional methods.2.In order to improve both segmentation accuracy and speed,a novel difficulty-aware prior-guided hierarchical network is proposed for the adaptive segmentation of breast tumors.A difficulty prior learning module is firstly proposed to learn characteristics of hard pixels that are segmented incorrectly due to complex characteristics.After that,a hard pixel processing unit is presented to learn more discriminative features for hard pixels.In the proposed network,a traditional feature extractor is used for an easy pixel,while a hard pixel processing unit is developed to deal with hard pixels more effectively.Therefore,the proposed network can segment breast tumors adaptively,further improving both segmentation accuracy and speed.The experimental results on three ultrasound image datasets show that the proposed method outperforms traditional deep learning methods and achieves a balance between accuracy and efficiency.3.In order to solve the problem that a single model fails to learn enough useful information about complex tumor regions,multi-level discriminative features embedding level set group is proposed for breast tumor segmentation.The level set group consists of several discriminative base segmentation models.In each discriminative base segmentation model,the discriminative features learned from the deep learning model are embedded in the level set method,which is helpful to improve the segmentation accuracy.Finally,the segmentation results of several discriminative base segmentation models are fused to give the final segmentation output.The experimental results on three ultrasound image datasets demonstrate the effectiveness of the proposed method.