Clinical Graph-Mediated Distillation for Unpaired MRI-to-CFI Hypertension Prediction
This paper addresses hypertension screening from inexpensive retinal fundus images by distilling knowledge from high-fidelity brain MRI—without requiring paired acquisitions from the same patients. The proposed Clinical Graph-Mediated Distillation (CGMD) constructs a clinical similarity graph using shared biomarkers (age, labs, etc.) to bridge disjoint MRI and fundus cohorts, propagates MRI teacher embeddings over the graph to impute patient-specific targets for fundus patients, and trains a fundus student with supervised, prior, and relational distillation losses. The approach aims to capture subtle vascular signals in fundus images by leveraging MRI-derived markers of small-vessel disease.
CGMD provides a principled mechanism for cross-modal knowledge transfer across disjoint cohorts, showing consistent improvements over standard distillation baselines in 5-fold cross-validation. However, the evaluation relies on very small cohorts (n=295 MRI, n=112 fundus) with high variance, limiting confidence in the reported AUC gains and raising concerns about generalization.
The graph-smoothing formulation (Eq. 2) and label-gated imputation (Eq. 3-4) are clinically grounded and well-motivated for denoising teacher embeddings and enforcing class-consistent transfer. Ablations confirm that graph smoothing is the critical component that enables effective relational distillation, and the kNN construction in biomarker space provides a plausible proxy for patient similarity when imaging pairs are unavailable.
The sample size is critically small for deep learning (ResNet-34/18 on ~100-300 patients), evidenced by large standard deviations (e.g., AUC 0.855 ± 0.127). With only 112 fundus patients, 5-fold CV yields ~22 validation patients per fold, making AUC estimates unreliable and confidence intervals wide. The paper omits a biomarker-only baseline (using the 16 shared clinical variables without any imaging), making it impossible to assess whether the MRI→fundus distillation adds value beyond the clinical data already shared between cohorts. Additionally, label gating requires ground-truth labels to construct imputed targets, limiting applicability to semi-supervised scenarios not discussed here.
Comparisons to adapted KD baselines using label-matched surrogate pairing are fair for the disjoint-cohort setting. However, the absence of statistical significance testing (p-values) is concerning given overlapping confidence intervals (e.g., KD AUC 0.803±0.101 vs CGMD 0.855±0.127). The lack of a held-out test set and reliance on 5-fold CV alone increases risk of optimistic estimates, particularly with such small cohorts. Figure 2(a) suggests performance improves with more data modalities, but does not isolate the contribution of MRI distillation from simple biomarker concatenation.
Code is available (https://github.com/DillanImans/CGMD-unpaired-distillation) but the dataset is private (tertiary medical center), preventing independent verification. Implementation details are thorough (PyTorch, ResNet architectures, 5-fold patient-level stratified splits), and hyperparameters are reported (λ_cls=λ_prior=λ_rel=1, σ=1, α=0.9, k=20 for MRI, k=5 for fundus). However, the criteria for selecting these hyperparameters (grid search, validation tuning) are not stated, raising questions about selection bias given the small sample size.
Retinal fundus imaging enables low-cost and scalable hypertension (HTN) screening, but HTN-related retinal cues are subtle, yielding high-variance predictions. Brain MRI provides stronger vascular and small-vessel-disease markers of HTN, yet it is expensive and rarely acquired alongside fundus images, resulting in modality-siloed datasets with disjoint MRI and fundus cohorts. We study this unpaired MRI-fundus regime and introduce Clinical Graph-Mediated Distillation (CGMD), a framework that transfers MRI-derived HTN knowledge to a fundus model without paired multimodal data. CGMD leverages shared structured biomarkers as a bridge by constructing a clinical similarity kNN graph spanning both cohorts. We train an MRI teacher, propagate its representations over the graph, and impute brain-informed representation targets for fundus patients. A fundus student is then trained with a joint objective combining HTN supervision, target distillation, and relational distillation. Experiments on our newly collected unpaired MRI-fundus-biomarker dataset show that CGMD consistently improves fundus-based HTN prediction over standard distillation and non-graph imputation baselines, with ablations confirming the importance of clinically grounded graph connectivity. Code is available at https://github.com/DillanImans/CGMD-unpaired-distillation.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.