Ensemble Deep Learning for Early Detection of Alzheimer's Disease Using MRI Brain Images: A Multi-Architecture Comparative Study

Abstract

Background: Structural MRI is difficult to diagnose early due to subtle changes at the onset of the disease, and MRI acquisition protocols have differences in their approach to early detection of the disease, which is the primary cause of dementia, namely, the Alzheimer disease (AD). This paper assesses the possibility of using a combination of complementary CNN backbones to enhance the performance of four-class AD staging.

Methods: we utilized a publicly available Kaggle dataset of 6400 axial T1weighted MRI images of four classes (No Impairment, Very Mild Impairment, Mild Impairment, Moderate Impairment). Two-stage pipeline used non-parametric localization, bias-field correction, and contrast enhancement (CLAHE), and then ResNet-50, EfficientNetV2-S and ConvNeXt-Base were used. The averaging of SoftMax output was done to make predictions. Data were divided as follows, training (70%), validation (20%), and test (10%) sets.

Results: EfficientNetV2-S was found to be 98.8% accurate (sensitivity 0.99, specificity 0.99, F1-score 0.98), ResNet-50 was found to be 97.5% accurate (sensitivity 0.98, specificity 0.98, F1-score 0.97), and ConvNeXt-Base was found to be 91.0% accurate (sensitivity 0.90, specificity 0.90 The 3-model ensemble mean-based model got an accuracy of 98.1% and sensitivity and specificity of above 0.97 in the four classes.

Conclusion: Localization and ensemble averaging by preprocessing made both multi-class AD staging on MRI more robust. The best overall balance was provided by EfficientNetV2-S, the highest precision in advanced stages confirmation was given by ResNet-50, and ConvNeXt-Base enhanced sensitivity in early-stage patterns. Clinical deployment needs to be externally multi-institutional validated.