MAUNext: a lightweight segmentation network for medical images.

Researchers

Journal

Modalities

Models

Abstract

In medical image segmentation tasks, accuracy and number of parameters are crucial factors for deep learning techniques to facilitate clinical research. By designing the lightweight backbone and the skip connection based on multiscale, channel attention machine and other strategies, a codec-based MAUNext approach has been developed. Firstly, a Multi-scale Attentional Convolution (MAC) module is introduced as the basic component of backbone encoder and decoder, resulting in higher accuracy with fewer parameters. Secondly, a collaborative Neighbourhood-attention MLP (NMLP) encoding module is designed to extract deep and abstract features, cooperating the MAC to further improve segmentation performance and lighten the network. Lastly, a tiny skip-connected Cross-Layer Semantic Fusion (CSF) module is proposed to bridge the semantic gap between encoder and decoder with minimal additional parameters. The MAUNext is evaluated extensively with eight state-of-the-art methods on three well-known datasets: 1) Kagglelung, 2) ISIC and 3) Brain. Experimental results demonstrate that our method is superior in terms of parameter numbers and accuracy when compared to other methods, presenting a promising solution for medical image segmentation tasks.© 2023 Institute of Physics and Engineering in Medicine.