Abstract
Deep Neural Networks (DNNs) often struggle with small-scale medical datasets due to overfitting and a lack of inherent spatial structure in tabular biomarkers. This paper introduces the COVLET (Covariance-based Wavelet-like) transform, a parametric framework that learns optimal multi-scale representations of neuroimaging data. By training on the scales of band-pass filters in a covariance-defined dual space, the model identifies discriminative biological patterns with high parameter efficiency. Evaluated on the ADNI dataset, our approach demonstrates superior classification accuracy for Alzheimer’s Disease stages compared to traditional machine learning and deep learning baselines.
Introduction
Deep learning models typically require massive datasets to generalize effectively. In clinical settings, data acquisition is often costly and restricted, leading to "small-data" regimes where complex models act as uninterpretable black boxes and are prone to overfitting. Furthermore, many neuroimaging biomarkers, such as cortical thickness, are represented as tabular data lacking the natural spatial or temporal order required for standard convolutional operations. This work addresses these limitations by introducing a physically interpretable model that identifies the most relevant "scales" of brain connectivity for disease classification.
Methods: The COVLET Framework
The core innovation is the COVLET transform, which adapts wavelet-like multi-scale analysis to the covariance structure of the input features.
The framework follows a structured pipeline:
- Covariance Mapping: The model begins by calculating the covariance matrix of features (e.g., regional brain measurements) to capture underlying biological relationships.
- Dual Space Projection: Data is projected into a "dual space" using the eigenvectors of the covariance matrix, serving as an analogue to frequency space in traditional signal processing.
- Trainable Scales: Instead of learning arbitrary filter weights, the model learns the optimal scales ($s$) of band-pass filters $g(s\Lambda)$. This allows the network to determine the specific resolution at which brain connectivity is most relevant for diagnosis.
- Classification: These multi-scale representations are concatenated and fed into a simple fully connected layer for final classification.
By backpropagating errors directly to the filter scales, the model identifies which "frequencies" of brain connectivity are most relevant for distinguishing Alzheimer’s patients from healthy controls.
Experimental Results
The study utilized data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), focusing on cortical thickness (MRI) and Tau protein accumulation (PET). The classification tasks included binary classification (Cognitively Normal vs. Early Mild Cognitive Impairment) and a complex 4-way classification (CN vs. EMCI vs. LMCI vs. AD).
As shown in the results table, the COVLET-based model consistently outperformed standard baselines. For the difficult task of distinguishing healthy controls (CN) from early-stage impairment (EMCI) using cortical thickness, the model achieved 85.8% accuracy, significantly higher than the ~74% achieved by Linear SVM and standard MLPs. In the 4-way classification, the model reached 80% accuracy, demonstrating robustness in identifying progressive disease stages. Notably, the model achieved these results with significantly fewer parameters than traditional MLPs, facilitating faster convergence and reducing the risk of overfitting.
Interpretability and Conclusion
Beyond performance, the model offers biological interpretability. Using Grad-CAM visualizations, the researchers confirmed that the model prioritizes regions such as the hippocampus, amygdala, and thalamus—areas well-documented in medical literature as primary sites of Alzheimer's-related atrophy—without being explicitly programmed to do so.
While the model shows great promise, future work will address the "curse of dimensionality" associated with high scale counts and explore multi-modal integration of MRI and PET data. In conclusion, the COVLET framework provides a powerful, interpretable alternative for neuroimaging classification in data-constrained environments.
Summary updated