Detection Of Abdominal Hernia With Lightweight Mesh Based On Deep Convolution Neural Network

Abdominal wall hernias are a common complication of abdominal surgery that cannot be cured spontaneously and can only be treated by implanting a mesh.In China,there are between 3 and 5 million abdominal wall hernia patients each year,with the incidence in elderly patients even reaching 38%,higher than the incidence of any other malignancy.Currently,the majority of clinically implanted meshes are weighted meshes,which are prone to chronic pain and have a high recurrence rate;lightweight meshes with sparse mesh and a lighter texture have gradually replaced weighted meshes.However,as an implanted foreign body,the lightweight mesh can still cause complications such as recurrence of abdominal wall hernia due to displacement and crumpling.The ABUS Automated 3-D Breast Ultrasound has been successfully used to detect not only breast tumours,but also lightweight meshes for abdominal wall hernias,overcoming the shortcomings of traditional methods.However,the volume of data generated by ABUS is very high.However,the volume of data generated by ABUS is very large,and manual review is tedious and time-consuming,making it prone to misdiagnosis and missed diagnoses.To address these issues,this paper proposes a fully automated detection localisation and segmentation method using deep learning by constructing a self-constructed ABUS ultrasound image dataset of abdominal wall hernia meshes.The main research work consists of two parts.(1)Considering the difficulties in localisation caused by defects such as noise and artefacts in ultrasound images,this paper proposes a CNN-based detection and localisation algorithm for ultrasound images of abdominal wall hernia meshes to help clinicians improve the accuracy and speed of mesh localisation.Using YOLOv3 as the basic framework,a spatial pyramid pooling structure is introduced in order to achieve the interplay of local detailed features and global semantic features;and migration learning is used to improve the robustness of the network and reduce overfitting.Through experiments,the average accuracy(Mean average precision,m AP)of the algorithm in this paper reached 90.15%,and the image recognition speed was 33.2FPS/S,which can effectively help doctors to detect and locate speckle areas quickly.(2)This paper uses U-Net as the basic framework,replaces the underlying 3×3convolution with residual block(Res-block)which has a stronger feature extraction capability,and adds the Convolutional Block Attention Module(CBAM)to improve the The target region saliency is improved by adding the Convolutional Block Attention Module(CBAM).Meanwhile,the ASPP(Atrous Spatial Pyramid Pooling)module combined with the null convolution module is used to effectively fuse multi-scale information while increasing the perceptual field to segment the ABUS images of ventral hernia meshes.Finally,in this work,the enhanced U-Net network can accurately segment the mesh contours with DSC,HD,PA and MIOU indices of 83.6%,0.42%,87.43% and84.97%,respectively.This study provided more accurate reference results for the detection localization and segmentation of abdominal wall hernia meshes,reduced the workload of physicians,significantly improved the diagnostic efficiency,and was of great clinical value for the secondary diagnosis and treatment of abdominal wall hernia patients,providing a solid foundation for the subsequent 3D reconstruction of abdominal wall hernia meshes.