Deep Cascaded Fully Convolutional Networks for Medical Image Segmentation

Segmentation of regions of interest (ROIs) in medical images is an important step for image analysis in computer aided diagnosis (CAD) systems. Recently, deep learning methods based on fully convolutional networks (FCN) have achieved great success in segmentation tasks on natural images. This success is primarily attributed to the ability of the FCN to leverage large datasets to hierarchically learn the features that best correspond to the appearance as well as the semantics of the images. However, there is a scarcity of annotated medical image training data due to the large cost and complicated acquisition procedures. Consequently, without sufficient training data to cover all the variations in ROIs, FCN usually produces poor boundary definitions for ROIs having inhomogeneous textures and fuzzy edges. Further, FCN tends to overfit to the visual characteristics that are present in the dominant classes, e.g., benign lesions and therefore, perform poorly on the minor classes which usually having more complex visual characteristics. In this work, we propose a new method for ROI segmentation to overcome the limitations mentioned above. We refer our method as deep cascaded fully convolutional networks (DCFCN), where we introduce to embed either multiple FCNs or other networks such as convolutional neural networks (CNN) and generative adversarial networks (GAN), to learn complementary visual characteristics, which are difficult to be learned in a single network. The ability to learn from a different network not only increase the overall feature diversity of the training data but also optimizes the segmentation results of the fuzzy, imbalanced and inhomogeneous ROIs. Our experiments on eight well-established public datasets demonstrate that our method is more effective than a routine FCN and its variants, and other state-of-the-art methods