Research On Deep Learning-based Methods For COVID-19 Image Feature Detection

Since late 2019,the global spread of novel coronavirus disease(COVID-19)has posed a serious health threat to people around the world,and a critical step in treating infected patients is to effectively evaluate and test their chest medical images.However,processing large volumes of complex medical images places a huge burden on clinicians’ time and effort,and test results based on subjective analysis vary from individual to individual.Therefore,the use of AI deep learning technology to automatically detect the precise area of infection from medical images will provide more reliable support for judging the development of the disease and clinical diagnosis and treatment.However,deep learning techniques often have the following challenges in the task of detecting the characteristics of COVID-19 lesions:(1)Lack of publicly available image feature dataset with annotated information on COVID-19 infection regions.(2)The height of the infected area in the COVID-19 image at different stages of illness,the texture of the lesion features is complex,and the contrast between the boundary blurred and the normal tissue is low,which brings great challenges to the detection.(3)The cost of manual labeling of medical imaging lesion features is high,and the difficulty of data acquisition seriously restricts the application of large-scale integration algorithms.In view of the above problems,this paper conducts an in-depth study of the new coronary pneumonia image feature detection method based on deep learning,which mainly has three aspects:(1)According to the evolution process and law of pulmonary CXR imaging in the early,developmental and severe stages of new coronary pneumonia positive cases,the infection area was divided into three lesion characteristics: pulmonary consolidation,infiltrate opacity and ground-glass opacity,and a COVID-19 multi-target feature detection dataset was established.(2)Aiming at the problem that the characteristics of COVID-19 lesions are difficult to detect,an improved image feature detection method for novel coronary pneumonia based on YOLOv8 is proposed.The DFEB module with self-attention mechanism is added in the feature extraction stage,which improves the global modeling ability and reduces the computational complexity of processing highresolution medical images through the calculation mode and sliding window mechanism of local attention.In the feature fusion stage,a multi-branch FEM module is added,and multiple parallel hole convolutional layers with different sampling rates are used to capture the feature information under different receptive fields,which effectively enhances the expression ability of features.Grad-CAM heatmap visualization technology is introduced to provide interpretability for model decisionmaking.Experimental results show that there are obvious performance advantages compared with other mainstream object detection algorithms.(3)Aiming at the problem that COVID-19 feature marker data is difficult to obtain,a new coronary pneumonia image feature detection method based on semi-supervised learning is proposed.In order to obtain high-quality pseudo-labels,the Soft-Teacher mechanism is used to evaluate the reliability of candidate boxes in unlabeled data,so as to reduce the negative impact of category noise on pseudo-labels.The WBF algorithm is used to fuse the useful position information in pseudo-labels with different confidence levels to improve the positioning accuracy of pseudo-labels.The automatic data augmentation strategy is used to select a combination of strong and weak data enhancement that is more suitable for COVID-19 images to apply to unlabeled data to improve the robustness of the model.Experimental results show that the semisupervised model can make full use of a large number of unlabeled data to obtain better detection performance,and effectively reduce the dependence on labeled data.