PROJECT TITLE :

GREN Graph-Regularized Embedding Network for Weakly-Supervised Disease Localization in X-ray Images

ABSTRACT:

Finding diseases in chest X-ray images requires only a few careful annotations, which saves a significant amount of human labor. Recent research has attempted to solve this problem by developing novel weakly-supervised algorithms, such as multi-instance learning (MIL) and class activation maps (CAM). Despite these advances, the results of using these methods are frequently inaccurate or insufficient. One of the reasons for this is that the pathological implications that are hidden in the relationships between different anatomical regions within each image as well as the relationships between different images are not taken into account. In this article, we argue that the cross-region and cross-image relationship, as a form of contextual and compensating information, is an essential component in order to achieve more consistent and integral region results. We propose the use of the Graph Regularized Embedding Network, or GREN, to model the relationship. This network makes use of both the information contained within individual images and the information contained within other images to locate diseases on chest X-rays. GREN first segments the lung lobes with the assistance of a pre-trained U-Net, and then it models the intra-image relationship between the lung lobes by making use of an intra-image graph to compare the various regions. In the meantime, a comparison of multiple images is accomplished by modeling the relationship that exists between in-batch images using an inter-image graph. Comparing multiple regions and images in order to make a diagnosis is a process that is analogous to how a radiologist is trained to think and how they make decisions. We make use of the hash coding and the hamming distance to compute the graphs, which are then utilized as regularizers to make training the neural network easier. This is done so that the deep embedding layers of the neural network are able to keep the structural information that is important to the localization task. Our approach achieves the state-of-the-art result on the NIH chest X-ray dataset for the weakly-supervised disease localization task by using this method. Our codes are accessible online.