An improved DeepLabv3+algorithm for semantic segmentation of pulse diagrams
10.3969/j.issn.1005-202X.2025.09.006
- VernacularTitle:一种面向脉象图语义分割的改进DeepLabv3+算法
- Author:
Linrong MANG
1
;
Zhaoxue CHEN
1
Author Information
1. 上海理工大学健康科学与工程学院,上海 200093
- Publication Type:Journal Article
- Keywords:
pulse diagram;
semantic segmentation;
bottleneck series fusion multi-scale pyramid pooling module;
channel-spatial serial attention mechanism
- From:
Chinese Journal of Medical Physics
2025;42(9):1159-1168
- CountryChina
- Language:Chinese
-
Abstract:
There is limited research on semantic segmentation of pulse diagrams using deep learning models.Therefore,an improved DeepLabv3+algorithm specifically for semantic segmentation of pulse diagrams is proposed.DeepLabv3+is served as the basic framework model,with its backbone network replaced by MobileNetV2 in the coding structure.This modification effectively reduces model size and parameter redundancy.Moreover,an improved bottleneck series fusion multi-scale pyramid pooling module is constructed to effectively expand the receptive field,improve segmentation accuracy,and further reduce parameter redundancy.An improved channel-spatial serial attention mechanism is integrated into the decoding structure for improving model performance.Concurrently,the sampling multiple is reduced for mitigating feature information loss.The improved DeepLabv3+model for semantic segmentation of pulse diagrams demonstrates exceptional capability in simultaneously and quickly detecting and identifying pulse curves and multiple elements in pulse diagrams.With a segmentation speed of 98.7 frames/s,the model achieves excellent performance metrics including 95.35%MPA and 88.66%MIoU,while maintaining a compact size of only 16.8 MB.Notably,the visualized segmentation results confirmes the superior performance of the proposed model over other models.These results suggest that the proposed model is particularly well-suited for efficient and fast semantic segmentation of pulse diagrams.
