Source-linked AI summary
Layer-Wise Relevance Propagation for Explaining Deep Neural Network Decisions in MRI-Based Alzheimer's Disease Classification
Moritz Böhle, Fabian Eitel, Martin Weygandt, Kerstin Ritter
TL;DR
Neural-network decisions for MRI-based Alzheimer’s classification are difficult to interpret clinically. The paper applies layer-wise relevance propagation to CNNs trained on structural MRI, finding patient-specific AD relevance patterns, little HC relevance, and literature-consistent regional associations, while noting model and parameter sensitivity.
Problem
CNNs can classify AD from structural MRI, but their complex decisions are difficult to interpret in clinical practice.
Method
The study generates LRP voxel-wise heatmaps from a CNN trained on ADNI structural MRI and compares them with guided backpropagation and atlas-based metrics.
Results
LRP provides individual-specific relevance patterns, little AD relevance in healthy controls, and literature-consistent regional relevance, especially in temporal-lobe areas including the hippocampus.
Takeaways & Limitations
LRP may assist clinicians in explaining CNN decisions for MRI-based AD diagnosis.
Takeaways & Limitations
Heatmaps identify contributing voxels but cannot establish underlying causes, voxel interactions, or causal relationships.
Abstract
from arXiv · showhide
Deep neural networks have led to state-of-the-art results in many medical imaging tasks including Alzheimer's disease (AD) detection based on structural magnetic resonance imaging (MRI) data. However, the network decisions are often perceived as being highly non-transparent, making it difficult to apply these algorithms in clinical routine. In this study, we propose using layer-wise relevance propagation (LRP) to visualize convolutional neural network decisions for AD based on MRI data. Similarly to other visualization methods, LRP produces a heatmap in the input space indicating the importance/relevance of each voxel contributing to the final classification outcome. In contrast to susceptibility maps produced by guided backpropagation ("Which change in voxels would change the outcome most?"), the LRP method is able to directly highlight positive contributions to the network classification in the input space. In particular, we show that (1) the LRP method is very specific for individuals ("Why does this person have AD?") with high inter-patient variability, (2) there is very little relevance for AD in healthy controls and (3) areas that exhibit a lot of relevance correlate well with what is known from literature. To quantify the latter, we compute size-corrected metrics of the summed relevance per brain area, e.g., relevance density or relevance gain. Although these metrics produce very individual "fingerprints" of relevance patterns for AD patients, a lot of importance is put on areas in the temporal lobe including the hippocampus. After discussing several limitations such as sensitivity toward the underlying model and computation parameters, we conclude that LRP might have a high potential to assist clinicians in explaining neural network decisions for diagnosing AD (and potentially other diseases) based on structural MRI data.
1 INTRODUCTION
Structural MRI and CNNs offer promise for Alzheimer’s disease classification, but their complex decisions are difficult to interpret clinically. The study motivates LRP heatmaps as patient-specific explanations of voxel contributions.
- AD causes medial-temporal cell death and atrophy visible in structural MRI, motivating imaging-based disease classification.
- CNNs can extract hierarchical image features directly from raw or minimally processed data for disease classification.
- CNN decisions are difficult to interpret because their multilayer architectures contain hundreds of thousands of trained parameters.
- Patient-specific heatmaps indicate each voxel’s importance for an individual classification decision and can support human-guided investigation.
- LRP decomposes the AD output score into input contributions, unlike gradient-based maps that show susceptibility to input changes.
- The study applies LRP to CNN-based structural-MRI classification, compares it with guided backpropagation, and evaluates atlas-based relevance metrics.
2 MATERIALS AND METHODS
The study trains a CNN on ADNI structural MRI scans and propagates class relevance backward to create individual heatmaps. It compares LRP with guided backpropagation using atlas-based, size-corrected regional metrics.
- Data: The dataset contains 969 1.5T T1-weighted MPRAGE scans from 475 AD and 494 healthy-control observations.These scans represent 193 AD patients and 151 healthy controls, with up to three time points.
- Model: The CNN comprises four convolutional blocks followed by two fully connected layers for binary classification.
- LRP: LRP traces output relevance backward layer by layer while conserving total relevance across layers.
- LRP: The β-rule adjusts the balance between positive and inhibitory contributions; β = 0 displays only positive contributions in the heatmap.
- Comparison method: Guided backpropagation maps absolute input-output gradients while setting negative gradients to zero at rectification layers.
- Analysis: Heatmaps are analyzed both group-wise and individually, with regional importance quantified using sums, size-normalized relevance, and gain relative to healthy controls.
3 RESULTS
LRP produces more AD-specific and individually variable relevance patterns than guided backpropagation. Relevance is concentrated in temporal-lobe regions and, for β = 0, relates to hippocampal volume.
- 3.1 Average heatmap comparison: LRP heatmaps show substantially more relevance in AD patients than healthy controls, whereas guided-backpropagation maps are susceptible for both groups.
- 3.1 Average heatmap comparison: False positives show less LRP relevance than true positives in similar areas, while guided backpropagation produces high gradients for false classifications.
- 3.2 Atlas-based importance metrics: LRP regional importance separates AD patients from healthy controls more consistently than guided backpropagation, whose gradients can be higher in healthy-control regions.
- 3.2 Atlas-based importance metrics: Size-normalized relevance makes the AD–healthy-control distribution difference more apparent under LRP, with greater variance among AD patients.
- 3.2 Atlas-based importance metrics: LRP gain is highest in temporal-lobe areas including the transversal temporal gyrus, hippocampus, planum temporale, and amygdala.
- 3.2 Atlas-based importance metrics: Regional overlap is strongest for relevance sum, followed by relevance density, while gain is relatively unstable for widely separated β values.
- 3.3 Individual heatmaps - fingerprinting and neurobiological relevance: Two high-confidence AD patients display markedly different relevance fingerprints, with one emphasizing temporal regions and the other frontal and superior-temporal areas.
- 3.3 Individual heatmaps - fingerprinting and neurobiological relevance: −0.560, p < 10^-3: hippocampal LRP relevance correlates negatively with hippocampal volume, unlike guided-backpropagation scores.
4 DISCUSSION
The study finds that LRP heatmaps provide individualized, biologically plausible explanations of CNN-based AD classifications, while highlighting important methodological limitations. Their interpretation depends on the metric, model, parameters, and lack of ground-truth explanations.
- Findings: LRP heatmaps showed high individual specificity, little AD relevance in healthy controls, and relevance concentrated in regions consistent with AD literature.Relevant regions included medial temporal structures such as the hippocampus, amygdala, parahippocampal gyrus, and entorhinal cortex, alongside frontal and parietal areas.
- Implications: Individual LRP relevance patterns formed variable fingerprints that may support future investigation of disease stages or AD subtypes.The paper presents this as a future research direction rather than an established diagnostic capability.
- Limitations: The proposed atlas-based metrics provide additional ways to assess heatmap explanations, but the gain metric produced unstable regional ordering for large, distant β values.Regional ordering was relatively stable for relevance sum and density, while heatmap sparsity increased with β.
- Limitations: Heatmaps lack a definitive ground truth and identify contributing voxels without revealing underlying causes, voxel interactions, or causal relationships.The observed association between hippocampal volume and LRP relevance does not establish causation.
- Limitations: Interpretations are constrained by classifier quality and atlas-to-anatomy mismatch, especially in atrophied AD brains.A heatmap indicates where the specific network model sees evidence, while atlas assignments may not perfectly match an individual patient's anatomy.
5 CONCLUSION
The study concludes that LRP explains individual CNN decisions in MRI-based AD diagnosis and supports quantitative comparisons of relevance across groups, regions, and methods. Its relevance patterns align with clinical findings but vary strongly between AD patients, while broader validation remains necessary.
- LRP heatmaps provide individual AD relevance for particular subjects, unlike gradient-based susceptibility maps focused on input changes.
- The framework introduces relevance density and relevance gain to compare heatmaps quantitatively across groups, brain areas, and methods.
- The metrics correlate well with clinical findings in AD but vary strongly between patients, supporting case-by-case evaluation.
- The authors identify evaluation across different network architectures and brain atlases, plus studies in MCI patients and known biomarkers, as future work.
FUNDING
The study acknowledges financial and institutional support from the German Research Foundation, the Manfred and Ursula-Müller Stiftung, and Charité – Universitätsmedizin Berlin.
- The work received support from the German Research Foundation, the Manfred and Ursula-Müller Stiftung, and Charité – Universitätsmedizin Berlin.
Relevance sum LRP
Regional relevance analyses compare LRP with gradient-based susceptibility using raw, size-normalized, and healthy-control-referenced measures. The figures emphasize area-size effects, temporal-lobe relevance, patient-specific patterns, and sensitivity to β.
- Absolute sums of LRP relevance and GB susceptibility are shown for the top 25 brain areas, with AD and HC averages compared separately.
- The absolute-sum metric correlates with brain-area size, motivating size-normalized relevance density for regional comparison.
- Size-normalized relevance density compares AD and HC distributions and highlights two high-confidence AD patients selected for maximal cosine distance in relevance-density space.
- Relevance gain divides regional AD importance by the corresponding HC average and displays correctly classified TP and TN cases against an identity-gain line of 1.
- Patient-slice visualizations mark medial and lateral temporal, amygdalar, inferior-frontal, and frontal-pole regions using thresholds relative to the average AD heatmap.
- Hippocampal volume is compared with hippocampal LRP relevance and GB susceptibility for correctly classified AD patients, with false positives shown additionally.