Fully Convolutional Architectures for Multi-Class Segmentation in Chest Radiographs

The recent success of Deep Convolutional Neural Networks on image classification and recognition tasks has led to new applications in very diversifying contexts. One of these is medical imaging where scarcity and imbalance of training data has hindered rapid development of neural network related applications. This paper investigates and proposes neural network architectures within the context of automated segmentation of anatomical organs in chest radiographs, namely for lung, clavicles and heart. By relating prior class data distributions to the objective function sparsely represented structures are methodologically emphasized. Scarce training sets and data augmentation are encountered with aggressive data regularization. The problem of highly imbalanced target object appearance in the input data is solved by modifying the objective function. The models are trained and tested on the publicly available JSRT database consisting of 247 X-Ray images the ground-truth masks for which available in the SCR database. The networks have been trained in a multi-class setup with three target classes. Our best performing model trained with the negative Dice loss function was able to reach mean Jaccard overlap scores of 94.1\% for lungs, 86.6\% for heart and 88.7\% for clavicles in the multi-label setup, therefore, outperforming the best state-of-the art methods for heart and clavicle and human observer on lung and heart segmentation tasks.