Transfer Learning and Stacked Ensembles for Neurological Disorder Classification

To enhance the identification and categorization of Alzheimer's Disease (AD) across four stages—Very Mild Dementia, Moderate Dementia, Mild Dementia, and Non-Dementia (Healthy Subjects). Leveraging a Kaggle dataset comprising 3,382 MRI brain images, the proposed methodology integrates transfer learning with the Inception V3 convolutional neural network to extract high-dimensional features, followed by ensemble stacking of machine learning (ML) models, including Neural Networks (NN 100x100, NN 70x70), XGBoost, CatBoost, AdaBoost, and a meta-learner. The dataset is enlarged to 299x299 pixels. It undergoes 10-fold cross-validation to check its performance. The features are saved in *.csv format for use in machine learning. Performance is assessed using AUC, Correctness Accuracy (CA), F1-score, Precision, and Recall, revealing the Stacking model's standout performance with an AUC of 0.959, CA of 0.870, and balanced metrics of 0.871, alongside NN 100x100's leading AUC of 0.967 and CA of 0.863. While XGBoost (AUC 0.928, CA 0.775) and CatBoost (AUC 0.881, CA 0.704) show moderate success, AdaBoost lags with an AUC of 0.681 and CA of 0.568, highlighting challenges with imbalanced data, particularly for the underrepresented Moderate Dementia class (64 images). The hybrid approach is good at identifying complex patterns in AD. It can help with early diagnosis and treatment. Future efforts will aim to augment the dataset volume, enhance configurations for the model, try different structures, and combine Various types of data.