A 2-Level Stacking Ensemble Framework for Multi-Class Alzheimer’s Disease Diagnosis Using CNN Feature Extractors and SVM-XGBoost Classifiers

Diagnosing Alzheimer's disease early is one of the most pressing challenges in modern medicine, often relying on doctors to examine magnetic resonance imaging, or MRI, scans of the brain. These images reveal the physical structure of the brain, showing how different regions shrink or change as the disease progresses through stages ranging from very mild memory loss to severe dementia. While human experts can spot these changes, the task is difficult because the differences between early stages are subtle, and the brain's anatomy is complex. To help, scientists have turned to artificial intelligence, specifically a type of computer program known as a deep learning model. These programs act like digital microscopes, trained to recognize patterns in medical images that the human eye might miss. By combining the insights of several different computer models, researchers hope to create a system that is not only more accurate but also more reliable than any single program working alone.

In a recent study, researchers Shashi Kant Mourya and Ajit Kr. Singh Yadav from NERIST in India developed a sophisticated system designed to classify brain scans into four distinct categories: non-demented, very mildly demented, mildly demented, and moderately demented. Their approach, described as a two-level stacking ensemble, works by gathering the opinions of three different artificial intelligence experts and then having a fourth, more experienced judge make the final decision. The first three experts are specialized computer networks trained to look at the MRI slices. Two of these networks are well-known, established designs called ResNet152 and DenseNet121, which have been used in many previous studies. The third expert is a custom-built network created specifically for this project, named the Shashi-CNN, which was designed from scratch to focus on the unique textures found in brain scans.

Instead of letting these three networks vote on the answer, the researchers had them work together in a specific sequence. First, each network analyzed the brain images and extracted a detailed list of features, essentially translating the visual data into a long list of numbers that describe the brain's condition. These lists were then combined and simplified to remove any confusing noise, keeping only the most important information. This refined data was fed into a first-level classifier, a type of mathematical tool known as a Support Vector Machine, which made an initial prediction for each image. Crucially, to ensure the system did not simply memorize the answers, the researchers trained this first layer using a method where the model was tested on data it had not seen during its training phase.

The final step involved a second-level judge, an advanced algorithm called XGBoost, which acted as a meta-learner. This judge did not look at the raw brain images; instead, it looked at the confidence scores and predictions generated by the first three networks and the first-level classifier. By analyzing how these different experts agreed or disagreed, the XGBoost model learned to weigh their opinions, deciding which expert was most trustworthy for each specific type of brain scan. This layered approach allowed the system to catch errors that a single model might miss, effectively creating a team where the weaknesses of one were covered by the strengths of another.

The results of this study were remarkably high. When tested on a dataset of 6,400 brain MRI slices, the full system achieved an accuracy rate of nearly 99.97 percent. This means that out of every 10,000 scans, the system correctly identified the stage of dementia in 9,997 of them. The system performed almost perfectly across all five different testing rounds, showing a level of consistency that suggests it is not just lucky but truly understands the patterns of the disease. In comparison, the best single model used in the study, the custom Shashi-CNN, reached an accuracy of about 98.4 percent, while the other established models scored lower. The researchers confirmed that the improvement gained by using the full team of models was statistically significant, meaning the difference was real and not due to random chance.

To ensure the system was making decisions for the right reasons, the researchers used a visualization technique called Grad-CAM. This tool creates a heat map over the original brain image, highlighting the specific areas the computer was looking at when it made its diagnosis. The visualizations showed that the model focused on anatomically relevant areas, such as the hippocampus and cortical regions, which are known to shrink in Alzheimer's patients. It did not get distracted by background noise or irrelevant parts of the image. This focus on the correct biological structures gives doctors confidence that the system is learning about the disease itself, rather than just memorizing patterns in the data.

While the study achieved near-perfect results on the dataset it used, the authors noted that the system has not yet been tested on data from different hospitals or with different types of MRI machines. The dataset contained a mix of brain scans, but one category, representing moderately demented patients, was much smaller than the others, which could influence how the system learns. Despite these limitations, the study demonstrates that combining multiple deep learning models with a smart voting system can create a powerful tool for distinguishing between the subtle stages of Alzheimer's disease. The findings suggest that such a framework could serve as a highly reliable aid for early diagnosis, potentially helping to catch the disease when it is most treatable.