Hybrid Topological and Deep Feature Fusion for Accurate MRI-Based Alzheimer's Disease Severity Classification

Early and accurate diagnosis of Alzheimer’s disease (AD) remains a critical challenge in neuroimaging-based clinical decision support systems. In this work, we propose a novel hybrid deep learning framework that integrates Topological Data Analysis (TDA) with a DenseNet121 backbone for four-class Alzheimer’s disease classification using structural MRI data from the OASIS dataset. TDA is employed to capture complementary topological characteristics of brain structures that are often overlooked by conventional neural networks, while DenseNet121 efficiently learns hierarchical spatial features from MRI slices. The extracted deep and topological features are fused to enhance class separability across the four AD stages. Extensive experiments conducted on the OASIS-1 Kaggle MRI dataset demonstrate that the proposed TDA+DenseNet121 model significantly outperforms existing state-of-the-art approaches. The model achieves an accuracy of 99.93% and an AUC of 100%, surpassing recently published CNN-based, transfer learning, ensemble, and multi-scale architectures. These results confirm the effectiveness of incorporating topological insights into deep learning pipelines and highlight the potential of the proposed framework as a robust and highly accurate tool for automated Alzheimer’s disease diagnosis.

💡 Research Summary

Alzheimer’s disease (AD) remains a major cause of dementia, and accurate staging of its progression from structural magnetic resonance imaging (MRI) is essential for early intervention and treatment planning. This paper introduces a novel hybrid deep‑learning framework that fuses Topological Data Analysis (TDA) with a DenseNet121 backbone to classify four AD severity levels (non‑demented, very mild, mild, and moderate) using the OASIS‑1 Kaggle MRI dataset.

The TDA branch processes each 2‑D MRI slice by constructing a cubical sub‑level filtration over a sequence of intensity thresholds. Persistent homology is computed for dimensions 0 (connected components) and 1 (loops), yielding persistence diagrams that are transformed into Betti curves. The curves are discretized into 100 bins per dimension, producing a 200‑dimensional topological vector that captures global shape information invariant to translation, rotation, and small intensity perturbations. A multilayer perceptron (MLP) reduces this vector to a 128‑dimensional embedding (F_TDA).

In parallel, the DenseNet121 branch receives the same slice after normalizing it to 248 × 248 pixels and replicating the single‑channel image into three channels. Pre‑trained ImageNet weights are used, and the classification head is removed. Dense connectivity enables feature reuse, and global average pooling condenses the convolutional feature maps into a 64‑dimensional embedding (F_CNN) that encodes local texture and spatial patterns relevant to AD pathology.

The two embeddings are concatenated into a 192‑dimensional fused representation, passed through two fully‑connected layers with ReLU activations and a dropout of 0.2 to mitigate over‑fitting. A final softmax layer predicts the probability distribution over the four classes. The model is trained end‑to‑end with the Adam optimizer, categorical cross‑entropy loss, a batch size of 32, and 50 epochs. No data augmentation, learning‑rate scheduling, or class‑balancing techniques are employed; a simple 90 %/10 % train‑test split with a fixed random seed is used.

Experimental results are striking: the hybrid TDA + DenseNet121 model achieves 99.93 % overall accuracy and a perfect AUC of 100 % on the held‑out test set, surpassing previously reported CNN, transfer‑learning, ensemble, and multi‑scale approaches on the same dataset. Visualization of the topological features via 3‑D PCA shows clear separation of the four AD stages, confirming that Betti‑0 and Betti‑1 descriptors carry discriminative information. The authors argue that the topological branch supplies global structural cues that complement the local, intensity‑driven features learned by the CNN, leading to improved robustness and generalization even with limited data.

Key contributions include: (1) the design of a hybrid, interpretable architecture that merges persistent‑homology‑based descriptors with deep convolutional features; (2) a systematic pipeline for extracting fixed‑length Betti vectors from MRI slices; (3) state‑of‑the‑art performance on OASIS‑1 without extensive data augmentation; and (4) an analysis demonstrating that topological information can substantially boost classification in neuroimaging tasks.

Limitations are acknowledged: the current implementation processes 2‑D slices independently, thus not fully exploiting 3‑D volumetric continuity. Future work may extend the approach to 3‑D TDA, incorporate additional persistent‑homology statistics (e.g., lifetimes, persistence landscapes), or integrate transformer‑based self‑attention mechanisms to capture long‑range dependencies.

In summary, the paper provides compelling evidence that topological data analysis can serve as a powerful complement to deep learning in medical image classification, offering a path toward highly accurate, data‑efficient, and interpretable tools for Alzheimer’s disease diagnosis and staging.