Scholarlink

<p>Alzheimer’s disease (AD) is a progressive neurodegenerative disorder impairing human cognition, memory, and behavior. Traditional diagnostic methods often lack sensitivity and specificity. To address this, we introduce NeuroFusionNet, a novel diagnostic tool leveraging neuroimaging data in AD. This model integrated a custom Convolutional Neural Network (CNN) and a pre-trained VGG16 model to extract rich, multi-level features, thereby enhancing diagnostic accuracy. We implemented a comprehensive preprocessing pipeline including brain region segmentation to isolate regions of interest and eliminate noise. The NeuroFusionNet leveraged both custom and pre-trained models to facilitate extracting a rich set of features, from low to high level, enhancing the classification accuracy. Local Interpretable Model Agnostic Explanations (LIME) further improved the interpretability of predictions. The model was evaluated on 600 test samples (120 per class) spanning five cognitive categories: AD, Cognitively Normal (CN), Early Mild Cognitive Impairment (EMCI), Late Mild Cognitive Impairment (LMCI), and Mild Cognitive Impairment (MCI). NeuroFusionNet achieved an overall accuracy of 0.81 (95% CI 0.779–0.841, <i>p</i> &lt; 0.001). Per-class performance metrics were as follows: AD: Precision 0.90 (95% CI 0.85–0.95), Recall 0.78 (95% CI 0.72–0.84), F1-Score 0.84, CN: Precision 0.67 (95% CI 0.60–0.74), Recall 0.97 (95% CI 0.94–1.00), F1-Score 0.79, EMCI: Precision 0.90 (95% CI 0.84–0.96), Recall 0.82 (95% CI 0.76–0.88), F1-Score 0.86, LMCI: Precision 0.95 (95% CI 0.90–1.00), Recall 0.87 (95% CI 0.81–0.93), F1-Score 0.90, EMCI: Precision 0.71 (95% CI 0.64–0.78), Recall 0.61 (95% CI 0.53–0.69), F1-Score 0.65. Training and validation curves over 50 epochs indicated robust learning with minimal overfitting. The integration of CNNs and LIME provided a powerful tool for early AD detection, with the potential to significantly impact clinical practices and patient outcomes. NeuroFusionNet performed better compared to both traditional and pre-trained models and high precision in AD, EMCI, and LMCI, along with LIME for interpretability, highlighted its clinical relevance. With further validation, NeuroFusionNet holds the potential to enhance AD diagnosis and clinical decision-making, offering a reliable and interpretable tool for clinicians.</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

NeuroFusionNet: integrating custom convolutional neural network and pre-trained models for accurate and interpretable Alzheimer’s disease diagnosis

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Abraham Varghese,
  • Andres Ricaurte-Fajardo,
  • Gowda Parameshwara Prashanth,
  • Manesh Thankappan,
  • Balakrishnan Kannan

摘要

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder impairing human cognition, memory, and behavior. Traditional diagnostic methods often lack sensitivity and specificity. To address this, we introduce NeuroFusionNet, a novel diagnostic tool leveraging neuroimaging data in AD. This model integrated a custom Convolutional Neural Network (CNN) and a pre-trained VGG16 model to extract rich, multi-level features, thereby enhancing diagnostic accuracy. We implemented a comprehensive preprocessing pipeline including brain region segmentation to isolate regions of interest and eliminate noise. The NeuroFusionNet leveraged both custom and pre-trained models to facilitate extracting a rich set of features, from low to high level, enhancing the classification accuracy. Local Interpretable Model Agnostic Explanations (LIME) further improved the interpretability of predictions. The model was evaluated on 600 test samples (120 per class) spanning five cognitive categories: AD, Cognitively Normal (CN), Early Mild Cognitive Impairment (EMCI), Late Mild Cognitive Impairment (LMCI), and Mild Cognitive Impairment (MCI). NeuroFusionNet achieved an overall accuracy of 0.81 (95% CI 0.779–0.841, p < 0.001). Per-class performance metrics were as follows: AD: Precision 0.90 (95% CI 0.85–0.95), Recall 0.78 (95% CI 0.72–0.84), F1-Score 0.84, CN: Precision 0.67 (95% CI 0.60–0.74), Recall 0.97 (95% CI 0.94–1.00), F1-Score 0.79, EMCI: Precision 0.90 (95% CI 0.84–0.96), Recall 0.82 (95% CI 0.76–0.88), F1-Score 0.86, LMCI: Precision 0.95 (95% CI 0.90–1.00), Recall 0.87 (95% CI 0.81–0.93), F1-Score 0.90, EMCI: Precision 0.71 (95% CI 0.64–0.78), Recall 0.61 (95% CI 0.53–0.69), F1-Score 0.65. Training and validation curves over 50 epochs indicated robust learning with minimal overfitting. The integration of CNNs and LIME provided a powerful tool for early AD detection, with the potential to significantly impact clinical practices and patient outcomes. NeuroFusionNet performed better compared to both traditional and pre-trained models and high precision in AD, EMCI, and LMCI, along with LIME for interpretability, highlighted its clinical relevance. With further validation, NeuroFusionNet holds the potential to enhance AD diagnosis and clinical decision-making, offering a reliable and interpretable tool for clinicians.