Skip to Content
You are currently on the new version of our website. Access the old version .
SensorsSensors
  • Article
  • Open Access

9 January 2026

23 Pages

A New Encoding Architecture Based on Shift Multilayer Perceptron and Transformer for Medical Image Segmentation

,
,
and
1
College of Advanced Interdisciplinary Studies, Central South University of Forestry and Technology, Changsha 410004, China
2
School of Mathematics and Physics, Hunan Institute of Technology, Hengyang 421002, China
*
Author to whom correspondence should be addressed.

Abstract

Accurate medical image segmentation plays a crucial role in clinical diagnosis by precisely delineating diseased tissues and organs from various medical imaging modalities. However, existing segmentation methods often fail to effectively capture low-level structural details and exhibit inconsistencies in feature connection, which may compromise diagnostic reliability. To address these limitations, this study proposes a novel Multilayer Perceptron–Transformer encoding architecture that integrates the Shift Multilayer Perceptron and Transformer mechanisms. Specifically, a SENet-based Atrous Spatial Pyramid Pooling module is designed to extract multi-scale contextual representations, while the Shift MLP refines underlying spatial features. Moreover, a channel–feature aggregation attention module is introduced to strengthen information flow between the encoder and decoder layers. Experimental results on the Automatic Cardiac Diagnostic Challenge dataset show an average Dice Similarity Coefficient (DSC) of 87.01% (83.32% for the right ventricle, 90.90% for the left ventricle, and 86.83% for the myocardium). On the Synapse multi-organ segmentation dataset, the proposed model achieves an average DSC of 79.35% and a 95% Haus Dorff Distance of 20.07 mm. These results demonstrate that MPT effectively captures both local and global anatomical structures, providing reliable support for clinical diagnosis.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.