ESNet: An Efficient Segmentation Network for Medical Image Segmentation

The automatic medical image segmentation technology based on deep learning has achieved high accuracy, but there is a dilemma that the parameters are huge and difficult to clinical application. To solve this problem, we propose an Efficient Segmentation Network (ESNet) model that saves parameters and floating point of operations while ensuring segmentation accuracy. We design a powerful feature extraction module DME to construct a hierarchical encoding sub-network to enable our model to have efficient feature learning capabilities, and a feature reuse-based DCD module to design a hierarchical decoding sub-network to effectively utilize the features generated by encoding to refine segmentation. To overcome the semantic gap between encoding and decoding, a conversion module ABT is established based on the idea of dense connection to communicate the encoding sub-network and the decoding sub-network while compensating for the spatial information of the decoding sub-network input. We evaluated our model on two different medical segmentation datasets, the 2018 Lesion Boundary Segmentation challenge and the 2018 Data Science Bowl challenge, achieving the SOTA performance on both Dice and Miou.