Abstract
Development of deep learning models to assess the degree of cognitive impairment on magnetic resonance imaging (MRI) scans has high translational significance. Performance of such models is often affected by potential variabilities stemming from independent protocols for data generation, imaging equipment, radiology artifacts, and demographic distributional shifts. Domain generalization (DG) frameworks have the potential to overcome these issues by learning signal from one or more source domains that can be transferable to unseen target domains. We developed an approach that leverages model interpretability as a means to improve generalizability of classification models across multiple cohorts. Using MRI scans and clinical diagnosis obtained from four independent cohorts (Alzheimer's Disease Neuroimaging Initiative (ADNI, $n = 1,821$), the Framingham Heart Study (FHS, $n = 304$), the Australian Imaging Biomarkers and Lifestyle Study of Ageing (AIBL, $n = 661$), and the National Alzheimer's Coordinating Center (NACC, $n = 4,647$)), we trained a deep neural network that used model-identified regions of disease relevance to inform model training. We trained a classifier to distinguish persons with normal cognition (NC) from those with mild cognitive impairment (MCI) and Alzheimer's disease (AD) by aligning class-wise attention with a unified visual saliency prior computed offline per class over all training data. Our proposed method competes with state-of-the-art methods with improved correlation with postmortem histology, thus grounding our findings with gold standard evidence and paving a way towards validating DG frameworks.
Introduction
Deep learning models are notoriously sensitive to "domain shift." In neuroimaging, a model trained on MRI scans from one hospital often performs poorly when tested on scans from another due to differences in scanners, imaging protocols, and patient demographics. Instead of learning biological markers of Alzheimer’s disease (AD), models may inadvertently learn to identify site-specific artifacts. While standard Domain Generalization (DG) techniques like data augmentation or adversarial training are often "disease-agnostic," this study proposes a "disease-informed" framework that uses model interpretability to guide training toward biologically relevant features.
Methodology: Attention-Guided Training
The authors propose a two-stage computational framework designed to orient the model toward disease-specific information.
- Prior Generation (Offline): A baseline 3D convolutional neural network (CNN) is first trained on a large source dataset. SHAP (SHapley Additive exPlanations) is then used to identify voxels most critical for predicting Normal Cognition (NC), Mild Cognitive Impairment (MCI), and AD. These averaged saliency maps serve as disease-informed priors.
- Attention-Guided Training: A new model is trained with an integrated attention module. A "similarity loss" ($L_{sim}$) is introduced to penalize the model if its internal attention maps deviate from the pre-computed SHAP priors. This ensures the model focuses on regions like the hippocampus and cortex rather than scanner noise.
Experimental Setup and Datasets
The study utilized four major independent cohorts, totaling over 7,000 participants. The NACC dataset served as the source domain for training, while ADNI, AIBL, and FHS were used as target domains to test generalization.
Additionally, a unique validation set of 23 participants with both MRI scans and postmortem histology scores provided a "gold standard" for biological grounding.
Key Results and Biological Grounding
The proposed method outperformed state-of-the-art DG methods across all target cohorts. Notably, on the AIBL dataset, it achieved a 73.4% accuracy, representing a 7.3% improvement over the baseline.
Beyond classification performance, the model demonstrated superior biological accuracy. Correlation analysis with postmortem histology (amyloid $\beta$ deposits and neurofibrillary tangles) showed that the model's attention maps closely matched physical brain damage observed by pathologists. This confirms that the model is learning true disease markers rather than spurious correlations.
Discussion and Conclusion
This work demonstrates that interpretability can serve as a powerful regularizer during training, not just a post-hoc analysis tool. By forcing the neural network to "look" at the right places, the researchers achieved better generalization across diverse clinical sites. While the current approach depends on offline prior generation and is computationally intensive due to 3D MRI volumes, it represents a significant step toward trustworthy and biologically grounded AI in neurology.
Summary updated