Adaptive encoding and comprehensive attention decoding network for medical image segmentation

Abstract

Medical image segmentation involves partitioning different tissues or lesion areas within medical images. Achieving automatic segmentation can markedly improve efficiency and accuracy, which is significant for biomedical clinical diagnosis. With the rapid development of deep convolutional neural networks (DCNN), U-Net has been widely used in medical image segmentation due to its encoder-decoder structure and skip connection. However, it is still hard for U-Net to handle certain challenging cases. In this study, we propose an adaptive encoding and comprehensive attention decoding network (AA-Net), which is derived from U-Net to address the issues of the semantic gap as well as the loss of spatial information during convolutions. AA-Net takes into account the different characteristics of the encoder and decoder. In the encoder, we design a simple Adaptive Calibration Module (ACM) to improve the representation ability of candidate features. In the decoder, we introduce a Comprehensive Attention Feature Extraction (CAFE) module, which employs multiple attention mechanisms after feature fusion to alleviate the semantic gap. Benefiting from CAFE, AA-Net can better handle the challenging cases where the segmentation targets vary in position, size, and scale. Additionally, we suggest a weighted hybrid loss function for precise boundary segmentation. We validate the effectiveness of AA-Net and each component on three biomedical image datasets. The results demonstrate that our method outperforms state-of-the-art methods in different medical segmentation tasks, proving it is lightweight, efficient, and general.