Diffuse Lung Diseases Textures Classification Based On Deep Learning

Diffuse lung diseases refer to a kind of tissue disorder that widely spread in lung regions.Radiologists diagnose these tissues through computed tomography(CT)images,but there exists the risk of misdiagnosis due to large quantity of images for checking and the complexity of pulmonary textures.Therefore,it is crucial to develop high precision algorithm for pulmonary textures classification of diffuse lung diseases to assist radiologists during diagnosis procedure.Existing algorithms for pulmonary textures classification have not fully utilized geometry information embedded in pulmonary textures,the feature learning capability is limited,the classification accuracy still has promotion space.Aiming at the deficiencies of existing algorithms,two deep learning-based algorithms for pulmonary textures classification of diffuse lung diseases are proposed in this article.Firstly,consider to combine geometry information embedded in pulmonary textures during pulmonary textures feature learning procedure.Each pixel in original CT image patch is utilized to calculate Hessian matrices followed by extracting eigen values,which are utilized to precompute corresponding geometry information image patch.Original CT image patches and corresponding geometry information image patches are fed into the dual-branch residual network fusing appearance and geometry information,which can effectively deepen the network and improve feature learning capability while learning and fusing appearance and geometry information of pulmonary textures.The geometry information image patches can also be utilized as an auxiliary material in later textures confirmation procedure conducted by radiologists.Then,through deep research on pulmonary textures classification problem,the geometry information in different kinds of pulmonary textures exhibits scale differences,the effective fusion of different scales' learnt feature information in the model can also play the role of learning geometry information embedded in pulmonary textures.Therefore,the multiscale attention network,which is fed with original CT image patches,is proposed to learn feature information of different scales through an end-to-end deep residual network,the multiscale feature fusion module follows to fuse feature information of different scales effectively.Meanwhile,the quality of feature information of different scales is promoted with attention mechanism module,which can automatically select feature information that are beneficial for classification task and automatically suppress feature information that have weak relation with classification task at the same time.In addition,the interpretability of network is promoted through conducting module function visualization on multi-scale feature fusion module and attention mechanism module.In order to evaluate the efficiency of the proposed two algorithms for pulmonary textures classification of diffuse lung diseases,several CT images are collected in this paper to train the algorithms,both patch and whole lung image slice-based tests show high classification accuracy.