Research On Detection And Recognition Of Medical Images Based On Deep Learning

Medical imaging is an important tool for the screening,diagnosis,treatment and evaluation of clinical diseases.One of the most representative devices is Computed Tomography(CT),which,because of its high sensitivity to soft tissue organs,can clearly show the detailed features of the lesion area on an anatomical map background,thus obtaining high diagnostic value images.Conventional CT imaging relies entirely on the experience of the reviewing physician,however,CT can only capture cross-sectional images within the overall contour of the patient from a wide-angle view,which leads to time-consuming and labor-intensive search for the lesion area in the complex organ and tissue background,and suffers from high subjectivity,low efficiency and insufficient quantitative analysis.Computer-aided Detection(CAD)system based on deep learning can achieve the goal of automatic examination and intelligent diagnosis with its powerful feature extraction and data analysis capabilities,which can effectively relieve the pressure of doctors’ film reading.The detection and identification of lesions or injury areas is one of the most important tasks in CAD,however,the lesion areas in CT images are smaller in size and have weaker edge features than those in natural scenes,and the similarity with neighboring tissues is extremely high,which results in a high rate of missed detection and false detection in clinical examination.In this regard,this thesis investigates the detection and recognition tasks of two small target lesions in the complex background of CT images: pulmonary nodules and foreign bodies in the eyes,and designs a detection and recognition system that better meets the clinical needs by combining the characteristics of their lesion images,respectively:1.To address the problems of current lung nodule detection algorithms with high leakage rate,high false positive rate,complex process and computational redundancy,this thesis proposes a lung nodule detection and recognition model based on the fusion of internal convolution and coordinate attention features,which uses YOLOv5 as the benchmark network and directly outputs the localization and classification information of lung nodules through the end-to-end operation mechanism,showing stronger robustness compared with other algorithms.First,this thesis establishes a lung nodule lesion diagnosis and treatment dataset with complete diagnosis and treatment information and a full range of cases,which solves the problem of insufficient image resources in the public dataset.Then,two novel modules are designed and added on the basis of YOLOv5 feature extraction network:one is the Transformer module for enhancing the network’s ability to capture global features in CT image space to reduce the false positive rate;the other is the Involution module for reducing the overall network computation to speed up the inference speed.Finally,a novel cross-scale feature fusion structure is constructed,in which high-dimensional semantic features are weighted by importance level and low-dimensional localization features are fused with semantic features after coordinate attention enhancement,which helps to deeply mine the semantic information in the model to improve the classification of pulmonary nodules,while enhancing the utilization of low-dimensional localization features to improve the problem of missed detection of pulmonary nodules.Finally,in the comparison test on this dataset,the improved detection model achieves 97.9 in m AP@0.5 index,while96.1% in sensitivity,98.8% in accuracy,and 84% in specificity index,all of which are substantially improved compared with the comparison algorithm.2.To address the current problems of large localization error,low accuracy of manual labeling and slow diagnosis of foreign body injury detection methods in the eye,this thesis proposes a light-weight eye foreign body detection and identification model with fused feature pyramid.Firstly,a perfect eye foreign body injury diagnosis and treatment data set is established.Then,the inverse residual structure is redesigned on the basis of the lightweight detection network Mobile Vi Tv3,the NAM-Attention attention mechanism enhancement layer is introduced to enhance the attention to salient features,and the Si LU activation function is introduced to improve the computational accuracy of the deep network and accelerate the model derivation speed.Finally,the path aggregation feature pyramid network is used in order to fuse multi-scale features and realize the information interaction between features at different levels of the network to improve the accuracy of judging the type of foreign objects in the eyes.In this thesis,the SPPFCSPC module reconfigures the connection between the spatial pooling layer and the convolutional layer to reduce the computational overhead,and introduces the SPD-Conv and Conv Mix modules to capture the key features in the spatial and channel directions in the deep network to improve the detection performance of foreign objects in the eyes.While maintaining the overall lightweight,the proposed model enhances feature extraction through feature pyramid structure,optimizes the way of feature information transfer and network computation,and completes the lightweight improvement of the model while improving the detection and recognition accuracy.The final comparison test proves that the relevant parameter indexes of this model: m AP@0.5 index of 97.2,sensitivity of 98.0%,accuracy of 93.5%,and specificity index of 88%.The thesis has 40 figures,9 tables,and 108 references.