Chest X-Ray Image Classification with Deep Learning

Publication Type:
Issue Date:
Full metadata record
Computer-aided diagnosis (CAD) systems have been successfully helped to clinical diagnosis. This dissertation considers one essential task in CAD, the chest X-ray (CXR) image classification problem, with the deep learning technologies from the following three aspects. First, considering most diseases existing in CXRs usually happen in small, localized areas, we propose to localize the local discriminative regions and integrate the global and local cues into an attention guided convolution neural network (AG-CNN) to identify thorax diseases. AG-CNN consists of three branches (global, local, and fusion branches). The global branch learns the global features for classification. The local branch localizes the discriminative regions, which avoids noise and improves misalignment in the global branch. AG-CNN fuses the global and local features for diagnosis in a fusion branch. Second, due to the common and complex relationships of multiple diseases in CXRs, it is worth exploiting their correlations to help the diagnosis. This thesis will present a category-wise residual attention learning method to concentrate on learning the correlations of multiple diseases. It is expected to suppress the obstacles of irrelevant categories and strengthen the relevant features at the same time. Last, a robust and stable CXR image analysis system should be able to: 1) automatically focus on the disease-critical regions, which usually are of small sizes; 2) adaptively capture the intrinsic relationships among different disease features and utilize them to boost the multi-label disease recognition rates jointly. We introduce a discriminative feature learning framework, ConsultNet, to achieve those two purposes simultaneously. ConsultNet consists of a variational selective information bottleneck branch and a spatial-and-channel encoding branch. These two branches learn discriminative features collaboratively. In addition, each of the proposed methods is comprehensively verified and analysed by conducting various experiments.
Please use this identifier to cite or link to this item: