International Journal of Science, Engineering and TechnologyAuthors: Shaik Shameer Basha, Prof. B. Sathyanarayana
Abstract: Objective: Standard Deep Learning(DL) evaluations for Alzheimer’s Disease (AD) often prioritize raw accuracy over clinical safety and cross-institutional generalization. This research proposes the Alzheimer’s Disease Generalized Clinical Reliability System (AD-GCRS), a novel framework designed to evaluate model stability and clinical risk across heterogeneous environments. Methods: The AD-GCRS framework leverages four transfer learning architectures VGG16, Xception, ResNet50, and EfficientNetB0 to classify MRI image scans into four progressive stages of impairment. We introduce a specialized evaluation suite, Clinical Deviation Error (CDE) to penalize stage-skipping misclassifications, the Index of Model Stability (IMS), and the Correct Class Index (CCI) for cross-dataset validation between ADNI and OASIS repositories. Results: Internal validation achieved accuracies exceeding 97%. However, cross-dataset testing revealed a significant "generalization gap," where the Xception model emerged as the most robust architecture with a CCI of 0.9708 and a restricted Major Error Rate (MCR) of 23.96%. Furthermore, we successfully resolved the "Scaling Paradox" in EfficientNetB0. It can be possible through a custom Lambda layer, restoring its diagnostic capability. Conclusion: The AD-GCRS provides a transparent pathway. For deploying AI in clinical settings by quantifying not just if a model is wrong, but how safely it fails
DOI: https://doi.org/10.5281/zenodo.18721750
