An Efficient and Reliability-Oriented Deep Learning Framework for Alzheimer’s MRI Classification

Reliable Alzheimer’s disease (AD) magnetic resonance imaging (MRI) classification requires evaluation beyond accuracy alone. This study investigates a four-class image-level framework based on ImageNet-pretrained ResNet-18 using 6,400 public MRI images labelled Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented. Original images were assigned to seventy percent training, fifteen percent validation, and fifteen percent test partitions before augmentation. The selected no-CLAHE pipeline used training-only transformations, 128 by 128 pseudo-RGB inputs, channel normalization with mean and standard deviation of 0.5, a 512 to 256 to 4 classifier, class-weighted cross-entropy with label smoothing, AdamW optimization, and five fixed seeds. Mean accuracy was 98.38 percent with a standard deviation of 0.38 percent, balanced accuracy was 98.86 percent with a standard deviation of 0.40 percent, macro F1-score was 90.73 percent with a standard deviation of 1.79 percent, and macro-AUC was 0.9993 with a standard deviation of 0.0009. For seed 42, validation-fitted temperature scaling reduced negative log-likelihood, Brier score, and expected calibration error without changing predicted labels. The model contains 11.31 million trainable parameters and achieved 3.08 ms batch-1 inference on a Tesla T4. Grad-CAM supported qualitative inspection. The results establish a reproducible, confidence-aware, perturbation-evaluated, and resource-efficient image-level framework within the stated public-dataset protocol.

@article{A2026,
  title     = {An Efficient and Reliability-Oriented Deep Learning Framework for Alzheimer’s MRI Classification},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {9},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Yashwant A and Rajaprakash S},
  doi       = {10.14569/IJACSA.2026.0170994},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170994}
}