Scholarlink

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder, which needs early diagnosis for successful intervention. This paper describes a comparative study of deep learning frameworks for multi-class classification of Alzheimer’s disease stages using MRI scans. Four methods were considered: a custom CNN (Convolutional Neural Network), and three transfer learning architectures (VGG16, ResNet50,and DenseNet121). The data set was comprised of 6,400 MRI images which were divided into four groups: Mild Demented, Moderate Demented, NonDemented, and Very Mild Demented. Models were trained for five epochs and assessed with accuracy, precision, recall, F1-score,and confusion matrices. Experimental results show that the VGG16 has the highest validation accuracy (61.3%) and better than the custom CNN (57.2%), ResNet50 (50.4%) and DenseNet121 (57.0%). However, class imbalance and small sample size caused all models to have poor performance in the Moderate Demented class. The results demonstrate the efficiency of the transfer learning, especially VGG16, in classifying Alzheimer’s stages and stress the necessity of the prolonged training with convergence monitoring and improved management of minority classes. This work is part of the research to build automated computer-assisted diagnostic systems that can assist clinicians in early diagnosis of Alzheimer’s disease.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative Analysis of Custom CNN and Transfer Learning Models for Alzheimer’s Disease Stage Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shivam Srivastava,
  • Pratibha Maurya,
  • Pankaj Kumar

摘要

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder, which needs early diagnosis for successful intervention. This paper describes a comparative study of deep learning frameworks for multi-class classification of Alzheimer’s disease stages using MRI scans. Four methods were considered: a custom CNN (Convolutional Neural Network), and three transfer learning architectures (VGG16, ResNet50,and DenseNet121). The data set was comprised of 6,400 MRI images which were divided into four groups: Mild Demented, Moderate Demented, NonDemented, and Very Mild Demented. Models were trained for five epochs and assessed with accuracy, precision, recall, F1-score,and confusion matrices. Experimental results show that the VGG16 has the highest validation accuracy (61.3%) and better than the custom CNN (57.2%), ResNet50 (50.4%) and DenseNet121 (57.0%). However, class imbalance and small sample size caused all models to have poor performance in the Moderate Demented class. The results demonstrate the efficiency of the transfer learning, especially VGG16, in classifying Alzheimer’s stages and stress the necessity of the prolonged training with convergence monitoring and improved management of minority classes. This work is part of the research to build automated computer-assisted diagnostic systems that can assist clinicians in early diagnosis of Alzheimer’s disease.