Research On Abnormal Lung Segmentation And Classification In HRCT

Early diagnosis of lung disease is cucial for making treatment decisions,which could improve the prognosis of patients.With the development of computer aided diagnosis system,the automatic segmentation and classification of abnormal lung has become an important means to assist doctors in rapid diagnosis.Different from healthy lung,ab-normal lung presents high complexity in high-resolution CT images,which poses serious challenges to lung segmentation and classification techniques.Firstly,the lesion areas of abnormal lung are accompanied by the presence of a large number of high attenua-tion patterns,with the result that the lung intensity and shape in CT images will appear challenging vatiations.These lesion areas will be excluded if using traditional lung seg-mentation methods.Secondly,most effective classification methods for abnormal lung rely on deep learning.However,the labeled datas are usually insufficient in the existing abnormal lung datasets due to the high cost in obtaining medical image datasets,leading to the poor generalization ability of the lung classification models.Therefore,in view of the above challenges,this paper aim to dive into the methods of abnormal lung segmentation and classification in high-resolution CT images.Firstly,this paper proposes a novel method for automatic segmentation of lungs based on 3D thresholding for the problem that it is difficult to segment the lung in the presence of high attenuation patterns.We believe that the segmentation accuracy of thresholding-based methods mainly relies on the performance of thresholding method itself.Searching for an optimum threshold is crucial because the quality of the initial lungs will determine that of the final segmentation results.Our proposed 3D thresholding method takes into account the spatial morphological characteristics of lung,which searches for the balance between lung details and external noise.It can preserve maximally the lung details,es-pecially those of high attenuation abnormalities.In addition,the fine-grain noise around lung surface can be removed by our asymmetric morphological operation As many lung details as possible can be preserved while completely removing the fine-grain noise.Fi-nally,the binary lung mask is further post-processed using convex hull and morphological operations,the irregular lung boundary can be refined to generate the final segmentation results.In HUX database,the proposed method achieves a mean Jaccard of 97.35%,a mean HD of 11.49 mm and a mean Avg HD of 0.19 mm for overall ILD scans.Secondly,in order to solve the problem of poor classification ability using existing classification models,we propose a semi-supervised abnormal lung classification method based on class activation map mix,which adds more samples and enriches the features of datas to improve the performance of classification.The semi-supervised learning method is first adopted to train all the datas including labeled datas and unlabeled datas,so as to avoid over-fitting on the insufficient labeled abnormal lung dataset.For semi-supervised learning,we add the following strategies to the model:consistency regularization and entropy minimization,generating the high confident pseudo labels in unlabeled data.In addition,in order to enrich the features of datas,we propose a mixed sample data augmen-tation method based Cam Mix.The labeled and unlabeled datas are predicted respectively to generate the class activation map masks with local consistent and arbitrary shape.Then the masks are mixed to another samples,and the associated labels are also mixed accord-ing to the number of pixels in the mask,which making full use of the connections and differences between samples.The performance of our proposed method is evaluated by comparing with existing abnormal lung classification methods.In an open abnormal lung dataset HUG,our method get a maximum F_Avgincrease 4.14%.Besides,Cam Mix has a better performance than other existing mixed sample data augmentation methods under the same semi-supervised baseline.