Research On Detection And Segmentation Algorithms Of Pneumonia Region Based On Convolutional Neural Networks

The detection and segmentation of pneumonia regions in medical images using convolutional neural networks is an auxiliary diagnostic tool in the clinical treatment of pneumonia.However,due to the complex representation of pneumonia in chest X-ray images and CT images,the feature extraction network selection,overlap detection boxes suppression,and loss function design parts are in urgent need of improvement.In this study,the improved algorithms for pneumonia region detection in chest X-ray images and pneumonia region segmentation in CT images,respectively,are presented as follows.(1)An improved Retina Net object detection model is proposed for the pneumonia region detection task in chest X-ray images.(a)The non-maximum suppression algorithm in the Retina Net target detection model is improved,and the Fuzzy Non-Maximum Suppression(FNMS)algorithm is proposed to obtain more accurate detection results by weighted fusion of overlapping detection frames.(b)Introducing Res2 Net,a multi-scale feature extraction network,into the Retina Net target detection model to improve the feature extraction capability for pneumonia regions of different sizes.(c)this paper improves the model ensemble method by integrating the overlapping boxes before NMS and using the FNMS algorithm to integrate the detection boxes of both models to get better detection performance.The experimental results on the RSNA pneumonia detection dataset show that the improved Retina Net object detection model proposed in this study has a better performance compared to the original Retina Net model and the excellent detection model in the RSNA pneumonia region detection competition.(2)Improved nn U-Net segmentation model applied to CT image pneumonia segmentation task.(a)The R-Dcie loss function is proposed to address the problems of imbalance between pneumonia region and normal region,segmentation annotation noise,and absence of pneumonia region in some slices of pneumonia segmentation data in CT images.The Dice loss function solves the problem of imbalance between the pneumonia region and normal region by focusing on the pneumonia region only.The NR-Dice loss function reduces the numerator order of the Dice loss function to obtain a Dice loss function that is robust to noise.The proposed R-Dcie loss function reduces the order of both numerator and denominator,corrects the wrong gradient direction of NR-Dice in the absence of pneumonia region to improve the robustness.(b)The detection model is integrated on the basis of the segmentation model,and the final segmentation performance is improved through the intersection of the detection results and the segmentation results.On the Mos Med Data 3D CT image pneumonia region segmentation dataset,as shown in the experimental result,the R-Dice significantly improved the segmentation performance of 2D and 3D versions of nn U-Net.In addition,the method of integrating detection models in segmentation tasks has significantly improved the segmentation performance.