Thorax Disease Identification And Localization In Chest X-ray Images

Chest X-ray is one of the most common used and cost-effective examinations in clinical medicine for the diagnosis of chest diseases.The appearance of chest X-ray images suffers from inter-similarity,e.g.,the “Pneumonia” and “ Pleural effusion” presenting the increasing opacity of chest X-ray images.In addition,due to factors such as patient position and depth of inspiration,the quality of chest radiographs is uneven.This makes the interpretation and diagnosis of chest radiographs more difficult.In clinical practices,radiologists not only focus on which diseases exist in the image,but also pay more attention to discovering the location of the lesion area.In the chest X-ray images,the lesion area is relatively small,and the similarity between the images is high,which makes pathological analysis difficult.Besides,there are many types of thorax diseases and many of them often happen at the same time.Current researches still lack reliable methods for extracting critical features of lesion areas,and there are few methods to explore the internal relationships between multiple diseases.Studying disease identification and localization in chest X-ray images can not only provide technical supports for computer-aided diagnosis,but also can improve the efficiency of diagnosis,which are worth to clinical practice.Targeting at the chest X-ray images,this thesis engages in two main critical problems,i.e.,the lesion areas are common small and difficult to learn the corresponding features for further analysis task;the complex relationships existing in multiple diseases.The main contributions are in the following three folds.(1)Currently,most thorax disease identification methods based on deep learning mainly focus on learning image features on the entire images.Using global images for disease feature learning and recognition does not take into account that local lesions are often very small and are easily affected by other healthy tissue regions in the image.They also do not consider the problem of large amount of noise in the images.To address the above problems,this thesis proposes a method of combining global and local features for disease identification and localization in chest X-ray images under the weakly supervised framework.First,an attention guided convolution neural network is proposed to locate the local discrimative region of images.Thus,it could avoid the influence of the disease-irrelevant regions or the regions ffected by noise in the image.Combining global and local features could help the eep leaning system detect small lesion areas.Local features could focus on the ost discriminative regions,and global features can compensate for the missing information in local features.The experimental results on the Chest X-ray14 dataset emonstrate the effectiveness of the proposed method.(2)In real scenarios,it it not rare to find that one patient is suffering from more than ne disease.There might be intrinsic correlations between different pathologies.Besides,some pathologies are easy to occur at the same time or once one happens,another would happen with large probability.Considering this observation,this issertation explores the inherent relationships between multiple diseases from the erspective of image feature learning,and aims to improve the level of computeraided diagnosis by enhancing the dependence between multiple diseases.We design a multi-label feature learning method based on visual attention mechanism to earn the correlations between features.The visual attention based method strengthens the relationship between disease-related features,and eliminates the interference of irrelevant features.It is used to enhance the dependence of multiple disease eatures in the feature space.Compared with the methods that learn the distribution f diseases in the sample space,learning the correlations in the feature space could void the problems of in sufficient samples or samping imbalance.The experimental results on the Chest X-ray14 dataset illustrate the effectiveness of the proposed ethod.(3)Thorax disease identification and localization benefit from learning the diseaserelated features and improving the correlations in the feature space.This thesis makes a further improvement and proposes a discriminative feature learning ethod for multi-label thorax diseases identification and localization.On the one and,a variational selective information bottleneck(VSIB)constraint is proposed o locate the critical discrimative regions end-to-end.Inspired by the traditional ariational information bottleneck theory,we introduce a selective mechanism on he feature spatial and channel dimensions into the variational information bottleneck.It is used to eliminate the interference of most of the non-lesion areas in he image and learn as much as possible the image features related to the diseases.Incorporating the selective mechanism could further strengthen the network filter aluable information,and thus improving the accuracy of localization and disease dentification.On the one hand,through the spatial-and-channel encoding(SCE) f disease image features,the long-range dependencies between image features in these two dimensions are automatically enhanced.Finaly,we fuse the VSIB and SCE features for final diagnosis.Experimental results on the Chest X-ray14 image dataset demonstrate that the VSIB and SCE could collaboratively learn discriminative features and improve the performance of disease identification and localization.