Deep Learning-based CT Image Assisted Diagnosis Of Bronchiectasis

Bronchiectasis is a common chronic respiratory disease that seriously affects the quality of life of patients and brings a heavy socioeconomic burden.CT imaging is the main tool for physicians to diagnose bronchiectasis,and more details of the lungs can be obtained through CT imaging.However,due to the nature of CT imaging,it also adds more mechanical burden to the physician’s diagnostic work.With the development of artificial intelligence,target detection technology in the field of computer vision can be used to assist in the diagnosis of bronchial dilatation,to detect bronchial dilatation in a timely manner and to reduce the burden of physicians.In this paper,we investigate the detection of bronchial dilation based on deep learning,and apply the target detection algorithm to detect and classify the target of bronchial dilation,using an optimized convolutional neural network to achieve the localization of the lesion of bronchial dilation.In order to demonstrate the effectiveness of the proposed algorithm,the bronchial dilatation detection model was trained and tested using the dataset provided by the imaging physicians of the Second Hospital of Shandong University in China-Japan Friendship Hospital.After the experimental demonstration and the judgment of the imaging physicians,better results were obtained.Therefore,the method proposed in this paper has some practical value for the improvement of clinical medical CAD systems.The main work of this paper is as follows.（1）This paper combines the dynamic Otsu thresholding method and morphological methods for the task of lung parenchyma extraction.Firstly,the dynamic Otsu thresholding method was used to binarize the CT images,and the lung mask was obtained by extracting the connected area with the largest area using the connected area analysis method;the lung parenchyma was extracted to the lung parenchyma using the morphological method.（2）The mechanism of anchor generation in Mask R-CNN feature pyramid network was improved for the focal characteristics of bronchial dilatation to improve the relevance of the model.The anchor generation strategy suitable for bronchial dilatation lesions was selected,and the aspect ratio of anchor generation was set to [1:3,1:1,3:1],which improved the detection accuracy of bronchial dilatation lesions.（3）The generation of multiple anchors in Mask R-CNN is prone to duplication,leading to the problem of repeated suggestions or even missed detection.In this paper,the Soft-NMS algorithm is introduced to improve Mask R-CNN by introducing Soft-NMS,and the effectiveness is verified by experiments.