DM-Net: A Dual-Model Network for Automated Biomedical Image Diagnosis

Biomedical image segmentation is an essential task in the computer-aided diagnosis system. An encoder-decoder based on a shallow or deep convolutional neural network (DCNN) is an extensively used framework for biomedical image analysis. To study and rethink the effectiveness of compounding both the shallow and deep networks for the medical image segmentation task, we propose a dual-model CNN architecture, called DM-Net, for biomedical image segmentation. DM-Net is composed of a shallow CNN structure at its left, called L-Net and a deeper CNN structure at its right, named R-Net. The L-Net is proposed to encode low-level contextual information and the R-Net is presented to produce high-level semantic feature maps. Furthermore, a novel crossed-skip connection (CSC) strategy is proposed to transfer information between the two side networks mutually. Extensive experiments demonstrate that our method outperforms representative approaches on three public medical image datasets.