首页|期刊导航|电子科技|融合SimAM注意力与自监督对比学习的脑出血图像分类

融合SimAM注意力与自监督对比学习的脑出血图像分类OA

Intracerebral Hemorrhage Image Classification Fusing SimAM Attention with Self-Supervised Contrastive Learning

中文摘要英文摘要

脑出血(Intracerebral Hemorrhage,ICH)是脑卒中相关死亡的主要原因,严重影响患者的生存率和生活质量.针对深度学习在脑出血识别中的训练困难和迁移性不足等问题,文中提出了一种融合简易、无参数的注意力模块(Simple,Parameter-Free Attention Module,SimAM)和自监督对比学习的SimCoNet(Simple Attention Module and Self-Su-pervised Contrastive Learning Network)网络模型.通过在自然图像数据集进行有监督预训练并结合大规模无标签脑出血数据集进行自监督学习来获取良好的可迁移特征表示.引入SimAM注意力模块并通过迁移学习在小型脑出血数据集进行微调分类.实验结果表明,所提方法在分类任务中的准确率为97.9%,相较于未引入SimAM注意力模块的模型,其准确率提升了0.2百分点.所提方法突破了医学图像数据量小、标注困难的限制,为脑出血诊断及计算机辅助医学诊断提供了新思路.

ICH(Intracerebral Hemorrhage)is the main cause of stroke-related deaths,seriously affecting the survival rate and quality of life of patients.In view of the problems of difficult training and insufficient migration of deep learning in the recognition of cerebral hemorrhage,this study proposes a SimCoNet(Simple Attention Module and Self-Supervised Contrastive Learning Network)network model that integrates a SimAM(Simple,Parameter-Free Attention Module)and self-supervised contrastive learning.Good transferable feature representations are ob-tained through supervised pre-training on natural image datasets and self-supervised learning in combination with large-scale unlabeled cerebral hemorrhage datasets.Introduce the SimAM attention module and fine-tune the clas-sification on the small cerebral hemorrhage dataset through transfer learning.The experimental results show that the accuracy rate of the proposed method in the classification task is 97.9%,which is 0.2 percentage points higher than that of the model without introducing the SimAM attention module.The proposed method overcomes the limitations of small data volume and difficult annotation of medical images,providing a new idea for the diagnosis of cerebral hemorrhage and computer-aided medical diagnosis.

金运龙;李一鸣;宋燕;孔平

上海理工大学 光电信息与计算机工程学院,上海 200093上海健康医学院 医疗器械学院,上海 201318上海理工大学 光电信息与计算机工程学院,上海 200093上海健康医学院 文理教学部,上海 201318

信息技术与安全科学

监督学习迁移学习深度学习对比学习自监督学习注意力机制医学图像分类ResNet50

supervised learningtransfer learningdeep learningcontrastive learningself-supervised learningattention mechanismsmedical image classificationResNet50

《电子科技》 2026 (10)

78-86,9

国家自然科学基金(62073223)上海市自然科学基金(22ZR1443400) National Natural Science Foundation of China(62073223)Shanghai Natural Science Foundation(22ZR1443400)

10.16180/j.cnki.issn1007-7820.2026.10.009

评论