Research On Chest Abnormalities Detection Based On Improved R-CNN In X-ray Images

X-ray image imaging has low cost,low radiation dose,and reasonable sensitivity to various pathologies.Because of its high availability,it has become the "flow entry" of modern medical images.In recent years,deep learning technology has gradually become a natural choice in the field of medical image analysis.It can assist in the interpretation of medical images by designing and implementing different computer vision models,and provide doctors with a reliable second diagnostic opinion.As the most commonly used radiological examination,chest Xray images are the core auxiliary means for screening and diagnosing various heart and lung abnormalities.Using deep learning technology to assist the interpretation of chest X-ray images,especially to accurately detect abnormal categories and locations is a research direction with important research value,but this fine-grained auxiliary diagnosis task has certain challenges.On the one hand,due to the imaging characteristics of X-ray images,abnormal regions in images have fewer discriminative features than salient objects in natural images,making it very difficult for the detection model to accurately distinguish different pathological patterns,such as pneumonia and pleural effusions all show increased opacity on X-ray images.On the other hand,abnormalities contained in chest X-ray images tend to be variable in scale and contain small objects that are difficult to detect,such as lung nodules.In addition,abnormal data of different categories show obvious long-tailed distribution characteristics,which makes the multi-label balanced training of the model more difficult.In view of the above three problems,this paper proposes effective solutions.The main research contents are as follows.Aiming at the difficulty of discriminative feature extraction in chest X-ray images,a feature extraction network Aug-Res Ne St guided by channel dimension attention and spatial dimension attention is designed.Among them,channel-dimensional attention realizes the effective interaction of channel information by adaptively reorganizing the features of different channels,filtering redundant features and retaining important features.Spatial dimension attention uses deformable convolution to break the limitation of regular receptive field,and flexibly extracts context information around the object as a reasonable supplement to channel dimension attention.By using the dual attention guided mechanism in the feature extraction network,its ability to extract discriminative features is effectively improved,thereby increasing the average detection accuracy of the detection model in this paper for 14 types of anomalies by 1.7 percentage points.Aiming at the problem of variable abnormal scale in chest X-ray images,Aug FPN with enhanced feature pyramid structure is proposed as a feature fusion network.This structure solves the key problems in the original feature pyramid network,such as loss of top-level feature information,large semantic gap of features in different layers,and insufficient utilization of multi-scale features by proposing residual feature enhancement,consistency supervision,and soft Ro I selection modules.Experiments show that the introduction of this structure effectively improves the overall detection performance of the model in this paper,and the average detection accuracy of the model for 14 types of anomalies is increased from33% to 34.1%.Aiming at the characteristics of long-tailed distribution of abnormal data of different categories,Seesaw Loss is used as the classification loss,which introduces two key factors in the common classification loss Cross Entropy Loss.During the training process,the excess negative sample gradient imposed on the tail category by the head category in the long-tail distribution is dynamically reduced,and the effect of the relative balance of the positive and negative sample gradients is achieved,thereby avoiding the classification bias problem caused by the unbalanced category distribution.Experiments show that using Seesaw Loss as the classification loss can effectively improve and balance the detection performance of the model for different types of anomalies,and improve the performance of the model in this paper by 0.8 percentage points.By solving the above three key problems,experiments on the Vinbigdata dataset show that the proposed multi-label abnormalities detection model for chest X-ray images has good performance,and finally achieved an average detection accuracy of 36.2%,surpassing most of the existing popular two-stage detectors.