Image segmentation is an important research direction in medical image processing.Automatic segmentation of medical images plays an important role in the early diagnosis of human diseases.In various types of medical images,ultrasound images have low contrast,fuzzy edges and contain a lot of noise.It is difficult to achieve good segmentation results by using traditional image segmentation methods.Although the image segmentation model based on deep learning can reduce the influence of the above characteristics through feature extraction,the mainstream encoder-decoder model has complex structure and large amount of parameters,which focuses on improving the segmentation accuracy and ignores the running speed of the model.To deal with the above problems,this thesis carries out some research.The main contributions of our thesis are three-fold:(1)This thesis introduces a novel and efficient encoder-decoder network,called Lightweight Attention Encoder-DecoderNetwork(LAEDNet),for automatic ultrasound image segmentation.In contrast to previous encoder-decoder networks that involve complicated architecture with numerous parameters,our LAEDNet adopts lightweight version of EfficientNet as encoder.On the other hand,a Lightweight Residual Squeeze-and-Excitation(LRSE)block is employed in decoder.To achieve trade-off between segmentation accuracy and implementing efficiency,we also present a family of models,from light to heavy(denoted as LAEDNet-S,LAEDNet-M,and LAEDNet-L,respectively),with varying lightweight version of EfficientNet backbones.To evaluate LAEDNet,we have conducted extensive experiments on Brachial Plexus Dataset(BP),Breast Ultrasound Images Dataset(BUSI),and Head Circumference Ultrasound Images Dataset(HCUS),where ultrasound images are suffered from high noise,blurred borders and low contrast.The experiments show that,compared with U-Net and its variants,e.g.,M-Net,U-Net++ and Trans UNet,our LAEDNet achieves better results in terms of Dice Coefficient(DSC)and running speed.(2)The segmentation accuracy of medical images is extremely important to assist clinicians in diagnosis and treatment.In order to further improve the segmentation accuracy of LAEDNet on the basis of ensuring the real-time performance of the network,this thesis redesigns an efficient decoding module CAS-LRSE for decoding.CAS-LRSE cascades the LRSE modules and modifies the residual connection method on this basis to transfer information across modules,which is more conducive to the learning of the network,so as to improve the feature expression ability of the network decoder.In order to keep the model lightweight,networks of different complexity are designed according to the difficulty of the segmentation task,which are called En-LAEDNet and En-LAEDNet-Tiny respectively.The effectiveness of the proposed method is verified by experiments on BP,BUSI and HCUS.Compared with LAEDNet-M,the improved network not only ensures real-time performance but also improves segmentation accuracy.(3)In order to improve the En-LAEDNet decoder’s ability to select the features of the encoder,so as to ignore the useless features with more noise,this thesis proposes a CAGM module for the decoder.This module acts before the decoder fuses high and low-level features,and uses the advantage of high-level features with rich semantic features to guide the selection of low-level features,resulting in more discriminative features.Experiments are carried out on BP,BUSI and HCUS.The experiments verify the effectiveness of the decoding mechanism of selecting features through CAGM and then merging features through CAS-LRSE.The improved network G-LAEDNet improves segmentation precision on the basis of ensuring real-time segmentation. |