Identification And Localization Of Thoracic Disease In Chest Direct Radiography Base On Weakly Supervised Learning

As one of the three deadly diseases,thorax diseases has always been seriously threatened human life and health.With the explosive growth of medical imaging,the pressure on imaging doctors is increasing.Researchers are committed to researching the intelligent diagnosis system of thorax diseases for direct digital X-ray chest medical images,which are used to improve the diagnosis accuracy and efficiency of thorax diseases and reduce the pressure on doctors.Relying on the essential characteristics of adaptive learning data,and effectively circumventing the limitations of artificially extracted features,deep learning has shown excellent performance in numerous computer vision tasks such as classification,detection,and semantic segmentation.Research on thorax disease recognition based on deep learning also achieved good results.However,most of these studies focus on the recognition performance of thorax diseases,and do not take into account that in actual applications,it is not only necessary to classify the disease,but also to accurately locate the disease.However,due to the scarcity of medical image positioning labels,it is often impossible to design and implement disease diagnosis systems based on detection models.In order to solve the above problems,this paper proposes a deep learning thorax disease recognition and weakly supervised localization method based on spatial attention guidance.First,in order to locate thorax diseases with weakly supervised learning,this paper combines with the multi-instance learning idea and designs a Top T%spatial pooling layer that aggregates the maximum T%response values and the minimum T%response values in the saliency map to connect the feature extraction module with the classifier,so the model only needs image-level labels to locate thorax diseases.Secondly,in order to improve the recognition performance of thorax diseases,this paper proposes a weakly supervised region-of-interest proposal operation based on saliency maps,which extracts the target area of interest of the global model and sends it to the local model for further disease recognition.The fusion model fuses the features of the global model with the one of the local model to achieve joint optimization of the model.Finally,in order to make effective use of limited bounding box labels,this paper proposes an annotation-constrained localization loss function,combined with multilabel classification loss to jointly optimize the model,and further improves the model’s localization and recognition performance.Based on the authoritative,public chest radiographic image dataset ChestX-ray14,this paper conducts experiments to verify the performance of the proposed algorithm,and realizes the diagnosis of 14 thorax diseases.The comparison with the existing related work proves that the algorithm in this paper has obvious advantages,reaching the optimal average AUROC result of 0.8522.At the same time,a self-certification experiment of each module is set up to verify the effectiveness.