Scholarlink

<p>Alzheimer’s disease (AD), a leading cause of dementia, presents persistent challenges in accurate staging and diagnosis from neuroimaging data. This study investigates advanced deep learning approaches for classifying AD from brain MRI, focusing on both predictive performance and resource efficiency. We evaluate a suite of convolutional neural network (CNN) architectures—including VGG16, ResNet50V2, MobileNetV2, EfficientNetB0, a custom lightweight CNN — under diverse configurations of activation functions (Mish, GELU, Swish, PReLU) and optimizers (AdamW, AdaBelief, NovoGrad, Lion, and SGD with cosine decay). Lightweight Vision Transformers (ViTs), such as MobileViT, are also benchmarked for comparison. Ensemble strategies are explored through both soft-voting and trainable attention mechanisms. The attention-based ensemble achieved the highest classification performance with an F1-score of 0.9492, while an efficient CNN-VGG ensemble reduced model size and memory usage by over threefold with only a minor performance trade-off. The best lightweight transformer model, MobileViT-Small, achieved an F1-score of 0.927 with exceptionally low inference latency, making it suitable for edge deployment. These findings highlight the effectiveness of combining model optimization with ensembling to produce accurate, interpretable, and resource-conscious models for Alzheimer’s diagnosis from MRI, advancing the state of neuroimaging-based decision support systems.</p>

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

Classification of Alzheimer’s disease using advanced deep learning and ensemble techniques

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月9日

查看全文
  • Viraj C. Desai,
  • Sucharitha Shetty,
  • T. Sujithra,
  • T. Manoj

摘要

Alzheimer’s disease (AD), a leading cause of dementia, presents persistent challenges in accurate staging and diagnosis from neuroimaging data. This study investigates advanced deep learning approaches for classifying AD from brain MRI, focusing on both predictive performance and resource efficiency. We evaluate a suite of convolutional neural network (CNN) architectures—including VGG16, ResNet50V2, MobileNetV2, EfficientNetB0, a custom lightweight CNN — under diverse configurations of activation functions (Mish, GELU, Swish, PReLU) and optimizers (AdamW, AdaBelief, NovoGrad, Lion, and SGD with cosine decay). Lightweight Vision Transformers (ViTs), such as MobileViT, are also benchmarked for comparison. Ensemble strategies are explored through both soft-voting and trainable attention mechanisms. The attention-based ensemble achieved the highest classification performance with an F1-score of 0.9492, while an efficient CNN-VGG ensemble reduced model size and memory usage by over threefold with only a minor performance trade-off. The best lightweight transformer model, MobileViT-Small, achieved an F1-score of 0.927 with exceptionally low inference latency, making it suitable for edge deployment. These findings highlight the effectiveness of combining model optimization with ensembling to produce accurate, interpretable, and resource-conscious models for Alzheimer’s diagnosis from MRI, advancing the state of neuroimaging-based decision support systems.