Detection And Analysis Of ABUS Light-Weight Mesh Based On Convolutional Neural Network And Level Set

Incisional hernia is an abdominal wall defect that occurs after abdominal surgery.It is caused by factors such as wound infection,poor recovery,or excessive exercise.Once the defect is formed,it cannot be cured by itself,and it has an increasing trend,which significantly affects the patient’s quality of life.At present,surgically implanting a mesh at the abdominal wall defect to withstand the tension between the original abdominal wall defect region and surrounding tissue has become the standard treatment for incisional hernias.Based on the development trend that fewer foreign bodies are better in the body,hernia repair mesh gradually becomes lighter.However,due to intra-abdominal pressure or strenuous exercise,the internal mesh may be displaced,shrunk and curled,which may lead to mesh-related complications such as intestinal adhesions,intestinal obstruction,or hernia recurrence.Accurate visualization of implanted meshes is a prerequisite for these mesh-related complications to be preventable and treatable.Automated breast ultrasound(ABUS)is an effective imaging device for lightweight mesh,however,ABUS imaging suffers from speckle noise and low imaging quality with increasing scanning depth.Meanwhile,lightweight meshes with special grid structure are more susceptible to speckle noise interference,and the size and degree of aggregation of lightweight meshes in slices with different imaging depths are variable.These problems make manual screening of lightweight meshes in a large number of ABUS images prone to miss and falsely detect lightweight meshes near the fascia or with tiny structures,it is also difficult to know how the lightweight mesh shrinks and moves in the patient over time.To solve the above problems,a detection and analysis method of ABUS lightweight mesh based on convolutional neural network and level set is proposed in this paper.The main research work is as follows:(1)First,preprocess the ABUS image of the lightweight mesh;second,the lightweight convolutional neural network PP-YOLO Tiny is optimized,and then trained from scratch for obtaining location priors of lightweight meshes in ABUS images;last,the distance regularized level set evolution(DRLSE)is combined for further processing to obtain the precise contour of the lightweight mesh.(2)The 3D model of the lightweight mesh is reconstructed according to different ABUS imaging acquisition times and the transformed actual size.The shrinkage and displacement of the lightweight mesh over time were then quantitatively analyzed by the vertical projected area and centroid shift of the 3D model at different times.The experimental results show that the proposed method can effectively detect the location of the lightweight mesh in the ABUS image,and the AUC values under the ROC curve of the lightweight mesh detection at any scale are all higher than 0.94.In addition,when using the evaluation indicators of lightweight mesh size and aggregation degree,the mean values of DSC and HD of contour detection results were 91.3% and 8.7,respectively,which were generally better than the compared methods.In the case group including 5 patients,the mean shrinkage rate error and displacement error between the experimental results and the manual results were 1.85% and 0.01 cm,respectively.In the in vitro experiments including 4 cases of gelatin and 3 cases of pig abdominal mass,the average shrinkage rate error and average displacement error were 0.75%,0.05 cm and0.56%,0.02 cm,respectively.These results demonstrate the effectiveness and feasibility of the method proposed in this paper.