通讯作者: 万艳丽, wan.yanli@imicams.ac.cn

DOI：10.12201/bmr.202505.00029

A Study on Improved Swin Transformer for Alzheimers Disease MRI Image Classification

• 摘要：目的/背景 阿尔兹海默病(AD)的早期诊断对于提高患者的生活质量和治疗效果至关重要。随着医学影像技术的发展,MR图像成为了识别和诊断AD的重要工具。然而,如何提高MR图像分类的准确性仍然是一个挑战。方法为了解决这一问题,本研究提出了一种改进的Swin Transformer模型,结合了多分辨率特征融合模块(MRFF)。该方法旨在增强模型对不同尺度特征的识别能力,从而提高分类准确性。结果 实验结果表明,改进后的Swin Transformer + MRFF模型在OASIS1数据集上的表现显著优于基础模型。该模型的分类准确率达到87.26%,召回率为91.43%。与基础模型相比,改进模型的准确率提高了2.43个百分点,F1值从87.26%提高至91.42%。此外,MRFF模块能够有效捕捉图像的局部细节和全局结构,显著提升了对Mild Demented和NonDemented类别的识别效果。通过结合数据增强和扩充策略,解决了样本不足和类别不平衡的问题,进一步提升了模型性能。结论 研究结果表明,基于Swin Transformer的改进深度学习方法在阿尔兹海默病MR图像分类任务中具有卓越的应用前景。该方法为早期诊断提供了强有力的支持,能够有效提高阿尔兹海默病的诊断精度。

  关键词： SwinTransformer、多分辨率特征融合（MRFF）、MR医学图像分类、阿尔兹海默病。; ; ; 

   

  Abstract: Purpose/Significance: Early diagnosis of Alzheimers disease (AD) is crucial for improving patients quality of life and treatment outcomes. With the advancement of medical imaging technology, MRI images have become an important tool for identifying and diagnosing AD. However, improving the accuracy of MRI image classification remains a challenge. Methods: To address this issue, this study proposes an improved Swin Transformer model, combined with a Multi-Resolution Feature Fusion (MRFF) module. The method aims to enhance the models ability to recognize features at different scales, thereby improving classification accuracy. Results: Experimental results show that the improved Swin Transformer + MRFF model significantly outperforms the baseline model on the OASIS1 dataset. The model achieves a classification accuracy of 87.26% and a recall rate of 91.43%. Compared to the baseline model, the improved models accuracy increased by 2.43 percentage points, and the F1 score rose from 87.26% to 91.42%. Additionally, the MRFF module effectively captures both local details and global structures of the images, significantly improving the recognition of Mild Demented and Non-Demented categories. By incorporating data augmentation and expansion strategies, the issues of insufficient sample size and class imbalance were addressed, further enhancing the models performance. Conclusion: The results demonstrate that the improved deep learning method based on Swin Transformer has excellent potential for Alzheimers MRI image classification tasks. This approach provides strong support for early diagnosis and can effectively improve the diagnostic accuracy of Alzheimers disease.

  Key words: Swin Transformer; Multi-Resolution Feature Fusion (MRFF); MR Medical Image Classification; Alzheimers Disease.

  提交时间：2025-05-21

  版权声明：作者本人独立拥有该论文的版权，预印本系统仅拥有论文的永久保存权利。任何人未经允许不得重复使用。
• html

• 图表

• 郁玲. 定量脑电参数对阿尔茨海默病治疗方案调整的指导意义. 2025. doi: 10.12201/bmr.202501.00062

  李筱琳, 隋欣, 宋娟, 包亚男, 林宇, 满子腾, 程甜甜, 杨宏艳. 基于WGCNA鉴定阿尔茨海默病的衰老关键基因. 2024. doi: 10.12201/bmr.202407.00036

  陈佳琦, 向红英, 凌迎春. 脑电图定量参数、血清miR-146a-5p对阿尔茨海默病患者认知功能障碍的预测价值分析. 2024. doi: 10.12201/bmr.202412.00027

  宫小翠, 安新颖. 面向医学科技文献分类的语义特征增强研究*. 2024. doi: 10.12201/bmr.202411.00088

  刘晶, 朱彦, 李宗友, 林东飞, 刘丽红, 史冬云. SNOMED CT疾病分类及概念模型研究. 2021. doi: 10.12201/bmr.202110.00005

  张胜发, 罗葳, 马玉环, 张晓宇, 赵远志, 周伟. 面向开放应用的生物医学大数据分类研究. 2024. doi: 10.12201/bmr.202411.00082

  王颖, 王兵, 高建强, 刘贵勤. 融合OBE理念的“线上+线下混合式”数字图像处理课程教学设计与实施. 2021. doi: 10.12201/bmr.202108.00011

  段磊, 张翔, 张研. 按病种的疾病诊疗难度测度及应用研究. 2024. doi: 10.12201/bmr.202403.00007

  张臣, 陈辉, 曹丰, 王玥琪, 柯任. 基于深度学习的左心房CTA图像自动分割研究进展. 2025. doi: 10.12201/bmr.202503.00042

  赵从朴, 袁达, 朱溥珏, 周炯, 陈政, 彭华. 基于BERT的医疗安全事件智能分类研究与实践. 2023. doi: 10.12201/bmr.202312.00021

• 序号	提交日期	编号	操作
  1	2025-02-06	bmr.202505.00029V1	下载

• 公开评论  匿名评论  仅发给作者 提交评论