Prostate Segmentation with Encoder-Decoder Densely Connected Convolutional Network (Ed-Densenet)

Prostate cancer is a leading cause of mortality among men. Prostate segmentation of Magnetic Resonance (MR) images plays a critical role in treatment planning and image guided interventions. However, manual delineation of prostate is very time-consuming and subjects to large inter-observer variations. To deal with this problem, we proposed a novel Encoder-Decoder Densely Connected Convolutional Network (ED-DenseNet) to segment prostate region automatically. Our model consists of two interconnected pathways, a dense encoder pathway, which learns discriminative high-level image features and a dense decoder pathway, which predicts the final segmentation in the pixel level. Instead of using the convolutional network as the basic unit in the encoderdecoder framework, we utilize Densely Connected Convolutional Network (DenseNet) to preserve the maximum information flow among layers by a densely-connected mechanism. In addition, a novel loss function that jointly considers the encoder-decoder reconstruction error and the prediction error is proposed to optimize the feature learning and segmentation result. Our automatic segmentation result shows high agreement (DSC 87.14%) to the clinical segmentation results by experienced radiation oncologists. In addition, comparison with state-of-the-art methods shows that our ED-DenseNet model is superior in segmentation performance.