Fetal Limb Deformity Detection Based On Deep Learning And Ultrasound Images

Fetal limb deformity is a common congenital malformation in clinic.Prenatal ultrasound is one of the most commonly used methods for clinical diagnosis of fetal limb deformity.The loss and deformity of fetal limb function will cause great harm to the family.In order to reduce the incidence of fetal limb deformity,the state and hospitals strongly advocate prenatal screening for pregnant women and give targeted intervention.At present,few studies use deep learning-based object detection methods to detect malformations in fetal limb ultrasound images.This study explores and studies the detection method of limb deformity in fetal ultrasound images based on deep learning technology.The main research contents and contributions are as follows :Firstly,this thesis uses the collected two-dimensional and three-dimensional ultrasound images of fetal limbs to produce a mixed data set that meets the needs of this experiment by formulating the labeling methods and standards of fetal ultrasound image data.The data set is preprocessed,including geometric transformation,image sharpening,gamma contrast transformation and other data enhancement methods,which effectively improves the quantity and quality of the data set.Secondly,this study first constructed a detection model based on YOLOv5 algorithm.Through the effective combination of CSPDarkNet-53,SPP,FPN and PAN structures,a feature extraction and fusion network was constructed,and finally the detection of different types of limb malformations in fetal limb ultrasound images was realized.Experiments on two-dimensional and three-dimensional ultrasound image datasets prove the effectiveness of combining two-dimensional and three-dimensional image data.Finally,based on the YOLOv5 network,the loss function,non-maximum suppression method and positive and negative sample matching strategy of the model are improved.The results show that the combination of CIoULoss and multi-anchor matching strategy can effectively improve the overall detection effect of the model and make the positioning of the model output prediction box more accurate.Thirdly,in order to further improve the performance of the detection model,this study attempts to apply the Transformer structure to the target detection task in the field of medical images,and proposes the Swin-YOLO model by combining SwinTransformer and YOLOv5.The experimental results show that the detection algorithm combining Transformer and CNN can not only extract enough local features,but also extract more global features with the help of self-attention mechanism.And through the hierarchical structure in the Swin model,different levels of feature information can be better fused.Finally,compared with the experimental results of other models,it is proved that the detection algorithm combined with Transformer is superior to the detection algorithm using only convolutional neural network in accuracy and recall rate.The fetal limb deformity detection method based on deep learning and ultrasound images proposed in this study has high detection accuracy and can be used as an effective method to assist doctors in clinical diagnosis.