首页|基于ResNet与分离注意力机制的肺部超声图像分类系统设计

基于ResNet与分离注意力机制的肺部超声图像分类系统设计

扫码查看
原文链接
  • 万方数据
  • 维普
目的 为解决传统的深度学习模型在处理具有多样性图像质量和微妙病变区域差异的肺部超声图像方面性能不佳的问题,设计一种基于残差网络(Residual Network,ResNet)与分离注意力机制的肺部超声图像分类系统.方法 采用ResNet152作为基础模型,结合分离注意力机制,通过对肺部超声图像进行预处理、数据增强和标准化处理,以提高模型的特征提取和分类能力.模型首先通过ResNet152进行深度特征提取,随后在各层引入分离注意力机制,增强模型对重要图像特征的关注,从而提高分类性能.结果 实验结果表明,优化后模型与原始模型相比,分类准确度在A线、B线、胸腔积液和肺实变上分别提升了0.51％、0.95％、14.17％和6.29％.通过消融实验,当同时使用Mish函数和分离注意力机制时,混合模型达到了97.92％的准确度.结论 本文提出的融合ResNet与分离注意力机制的肺部超声图像分类系统模型可为临床超声诊断提供较高的参考价值.
Design of Lung Ultrasound Image Classification System Based on ResNet with Split Attention Mechanism
Objective To address the issue of poor performance of traditional deep learning models in processing lung ultrasound images with diverse image quality and subtle differences in lesion areas,to design a lung ultrasound image classification system based on residual network(ResNet)and separation attention mechanism.Methods Using ResNetl52 as the basic model,combined with the separation attention mechanism,the feature extraction and classification ability of the model were improved by preprocessing,data enhancement and standardization of lung ultrasound images.ResNetl52 was first used for deep feature extraction,and the separation attention mechanism was then introduced at each layer to enhance the model's attention to important image features,thus improving the classification performance.Results The experimental results showed that compared with the original model,the optimized model improved classification accuracy by 0.51％,0.95％,14.17％,and 6.29％on A-line,B-line,pleural effusion,and lung consolidation,respectively.Through ablation experiments,the mixed model achieved the highest accuracy of 97.92％when using both the Mish function and separation attention mechanism simultaneously.Conclusion The research results indicate that this paper proposed lung ultrasound image classification system model that integrates ResNet and separated attention mechanisms can provide high reference value for clinical ultrasound diagnosis.

residual network(ResNet)Mish functionResNetl52lung ultrasound imagesdepth feature extractionimage classificationultrasonic diagnosis

杨倩茹、郭峻氚

展开 >

新疆维吾尔自治区人民医院重症医学科,新疆乌鲁木齐 830000

残差网络 分离注意力机制 Mish函数 ResNet152 肺部超声图像 深度特征提取 图像分类 超声诊断

新疆维吾尔自治区自然科学基金新疆维吾尔自治区人民医院院内基金

2021D01C15620230138

2024

10.3969/j.issn.1674-1633.2024.10.009

中国医疗设备
中国整形美容协会

中国医疗设备

CSTPCD
影响因子：0.825
ISSN：1674-1633
年,卷(期)：2024.39(10)