Segmentation of Multiple Structures in Chest Radiographs Using Multi-task Fully Convolutional Networks

Segmentation of various structures from the chest radiograph is often performed as an initial step in computer-aided diagnosis/detection (CAD) systems. In this study, we implemented a multi-task fully convolutional network (FCN) to simultaneously segment multiple anatomical structures, namely the lung fields, the heart, and the clavicles, in standard posterior-anterior chest radiographs. This is done by adding multiple fully connected output nodes on top of a single FCN and using different objective functions for different structures, rather than training multiple FCNs or using a single FCN with a combined objective function for multiple classes. In our preliminary experiments, we found that the proposed multi-task FCN can not only reduce the training and running time compared to treating the multi-structure segmentation problems separately, but also help the deep neural network to converge faster and deliver better segmentation results on some challenging structures, like the clavicle. The proposed method was tested on a public database of 247 posterior–anterior chest radiograph and achieved comparable or higher accuracy on most of the structures when compared with the state-of-the-art segmentation methods.