acceptodds
Under review as a conference paper at ICLR 2027

Reliable Neuroimaging Model Interpretation via Closed-Form ROI Influence Posteriors

Abstract

Neuroimaging-based machine learning models are increasingly used to study brain disorders and aging, yet their interpretations are often reported as deterministic importance maps over brain regions. In region-of-interest (ROI) analyses of structural MRI and fMRI, such point explanations can be unreliable: limited sample sizes, heterogeneous cohorts, and training uncertainty can change which brain regions appear influential, while standard attribution methods rarely quantify whether a regional effect is stable. We propose Closed-Form Influence Posterior (CFIP), a probabilistic framework for reliable neuroimaging model interpretation. CFIP defines regional influence as the model-output change induced by intervening on an ROI, uses a Laplace approximation to model parameter uncertainty, and propagates this uncertainty through a first-order linearization of the ROI influence functional. Under this Laplace-linearized construction, the resulting closed-form posterior over ROI influences provides an analytic distributional view of regional effects, supporting uncertainty-aware interpretation, reliability assessment, and a decomposition of attribution instability into model uncertainty and data heterogeneity. We validate CFIP across multi-cohort neuroimaging benchmarks spanning structural MRI-derived regional features and fMRI connectivity representations, including Alzheimer’s disease and aging settings. CFIP delivers faithful regional interpretations, uncertainty estimates that track observed attribution variability, and efficient closed-form inference without costly sampling. These results support CFIP as a practical reliability layer for anatomically and functionally meaningful neuroimaging model interpretation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
AI for health, xAI

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RKmmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CFIP estimates uncertainity in ROI-level neuroimaging explanations. It defines an ROI's effect as the change in model logit after replacing that ROI. It applies a Laplace approximation to model parameters and derives a Gaussian posterior over ROI effects. Tests include ADNI, OASIS, HCP-Aging, synthetic data, and comparisons with ensembles and HMC.

Critical strengths
  • Influence posterior is simple to compute, and the exact last-layer case is clearly derived. Joint covariance across ROI effects is useful.
  • Evaluation goes beyond attribution maps: retraining stability, HMC comparisons, calibration, and runtime measurements.
Critical weaknesses
  • Methodological novelty seems limited for ICLR.
    • With fixed features and the last-layer assumption, this is a standard Gaussian pushforward of a Laplace posterior.
    • Posterior mean is exactly the deterministic ROI-removal effect.
    • So better deletion ranking compared with gradient/SHAP methods does not show a benefit from posterior uncertainty itself.
  • Reliability is inconsistent in harder settings. The cognitively normal to mild cognitive impairment task has 0.708 coverage for nominal 90% intervals in the main analysis, and Laplace intervals underestimate HMC variability.
  • Most experiments fix the feature extractor too, leaving out an important source of uncertainty.
Questions for authors
  • What does CFIP add over applying standard last-layer Laplace directly to deterministic ROI-removal effects? Please separate gains from the attribution definition and gains from uncertainty estimation.
  • Can the authors explain the undercoverage against retraining/HMC and show when all-layer uncertainty improves reliability in the clinical tasks?
3 comments

We thank Reviewer RKmm for the careful review. We appreciate the reviewer's recognition that the exact last-layer case is clearly derived and that the evaluation goes beyond attribution maps.

W1 + W3 — novelty; fixed feature extractor

Reply: Our contribution is the uncertainty estimation in brain neuroimaging tasks: a closed-form posterior over regional influence that needs no repeated sampling, for neuroimaging models whose explanations are otherwise reported as single point maps. The influence functional is not part of our contribution. The deletion-based functional is a simple starting point; the same construction applies to other influence definitions (Table 14 already applies it to a gradient functional), and we will explore better ones in future work. On the fixed feature extractor, using extracted regional features is standard practice in brain neuroimaging (Fischl, 2012; Cuingnet et al., 2011; Rathore et al., 2017), and training from scratch does not beat it: a reproducible ADNI benchmark showed no advantage for image-level CNNs over an SVM (Wen et al., 2020), and UK Biobank benchmarking gave deep models no edge over linear ones on brain scans (Schulz et al., 2020). CFIP is also not limited to a fixed extractor: the full-network tangent variant (Sec. 4.5) includes all layers in closed form.

Q1 — separating gains from the attribution definition and from uncertainty estimation

Reply: We separate the two by fixing one and changing the other. With the uncertainty fixed, the functional sets the ranking: under the same Laplace posterior, the deletion-based functional gives AOPC 0.177 against 0.127 for the gradient one (Table 14). With the functional fixed, the gain from our uncertainty estimation is reliability information at closed-form cost. It matches the ranking quality of a ten-model ensemble from a single trained model (AOPC 0.177 vs 0.172, Table 14; 0.056 vs 0.340 ms per subject, Table 7), it reproduces 4,000-draw posterior sampling to within 0.001 with no sampling, and it preserves the variance ordering of the HMC reference (ρ=1.000\rho = 1.000, Table 5). Its variances also track retraining variability, with rank agreement above 0.90 on all three ADNI tasks (Tables 3 and 18), which is what lets it flag unstable regions: ROIs that enter the top ten only intermittently across retrainings carry 1.24 times the uncertainty of those always selected (Sec. 5.6).

W2 + Q2 — under-coverage on harder tasks; all-layer version

Reply: The under-coverage appears on harder tasks such as CN–MCI because the signal is weaker, so different training runs settle on different solutions: ranking stability across retrains drops from 0.652 on CN–AD to 0.541 on CN–MCI (Table 12). A posterior built from one trained model does not include this between-run variation, so its intervals are narrower than the spread of retrained models. Against HMC, which uses the same trained model, the gap is small: Laplace keeps the variance ordering (ρ=1.000\rho = 1.000) and its variances are 92–95% of HMC’s (Table 5). We have therefore introduced a calibration step: the uncertainty hyperparameter is selected on a held-out set so that the 90% coverage target is met, and on held-out subjects this holds across all three ADNI tasks (table). All-layer uncertainty matters when the extractor itself varies across training runs: by Proposition 2 it adds the extractor’s contribution on top of the last-layer variance, and Table 2 shows this contribution grows with extractor depth. The last row reports it on the three clinical tasks. On the clinical tasks (last row) it removes the under-coverage on CN–MCI, raising coverage from 0.793 to 0.964, but it is conservative on all three tasks, with wider intervals.

Retraining Cov@90CN–ADAD–MCICN–MCI
Diagonal, nominal λ\lambda0.931 ± 0.0820.893 ± 0.1940.708 ± 0.171
Diagonal, coverage-calibrated0.903 ± 0.0300.908 ± 0.0240.903 ± 0.077
Full GGN, nominal λ\lambda0.924 ± 0.0510.912 ± 0.1450.793 ± 0.131
Full GGN, coverage-calibrated0.914 ± 0.0380.919 ± 0.0210.921 ± 0.059
All-layer, nominal λ\lambda0.983 ± 0.0430.983 ± 0.0340.964 ± 0.036

References: Wen et al., 2020 (https://arxiv.org/abs/1904.07773v6); Schulz et al., 2020 (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7447816/).

Dear Reviewer RKmm,

We hope we’ve addressed all your concerns. Since the discussion period ends in <6 hours, please let us know if you need any further clarification. If no further clarifications are needed, may we kindly request that you revise your scores upwards?

Thank you, The Authors

Thank you very much for the detailed response and the additional results. I appreciate authors' clarifications and the effort to address my concerns. A few points remain:

  • Methodological novelty: The authors clarify that the influence functional is not part of the contribution. The proposed closed-form posterior is still a standard Gaussian pushforward of a Laplace approximation. Its methodological novelty remains unclear.
  • Attribution versus uncertainty: The rebuttal separates these contributions more clearly. However, the posterior mean equals the deterministic ROI-removal effect, the reported gains in attribution faithfulness is not a benefit from uncertainty estimation itself.
  • Fixed feature extractor: The given references support the practical use of simpler neuroimaging models, but do not show that uncertainty in the learned feature extractor can be neglected.
  • Calibration and reliability: The new coverage results are nive, particularly for CN-MCI. The calibration protocol and the independence of calibration and evaluation data still need clarification. The all-layer approach also produces relatively conservative intervals.

Overall, the rebuttal partly addresses my reliability concerns, but the central concern about methodological novelty remains. I therefore maintain my original rating of 2 (Weak Rejection).

Reviewer_S1K5OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Closed-Form Influence Posterior (CFIP) for uncertainty-aware interpretation of neuroimaging predictors. We define an ROI's influence as the logit change after replacing that region with a specified baseline. CFIP propagates a Laplace-parameter approximation through this influence functional, yielding a Gaussian distribution that is exact for a fixed-feature affine output layer and approximate for a nonlinear network via a tangent construction. The resulting distribution provides regional sign probabilities, joint covariance, and a decomposition of model/data variance. Validation includes structural MRI and fMRI tasks, synthetic feature effects, Bayesian baselines, HMC comparisons, and checks of nonlinear propagation error.

Critical strengths
  1. The attribution target and the conditions for analytical inference are explicit. Equation 1 measures model reliance under a specified ROI replacement. Theorem 1 correctly states when its Gaussian distribution is exact under the parameter approximation, and the full-network construction distinguishes the tangent model from the original nonlinear network. This makes the quantities being interpreted and the assumptions behind their uncertainty clear.

  2. This paper examines approximation limits and provides relevant controls. The translated prior in Appendix C, the extractor contribution in Proposition 2, and the shallow-ReLU remainder analysis carefully account for the full-network extension. Tables 5 and 6 compare against HMC and nonlinear-network evaluations, while Table 14 separates attribution-functional choices from uncertainty estimators.

Critical weaknesses
  1. The distinction from sampling-based explanations overstates the methodological advance. Figure 1 describes sampling as producing "moments only" and lacking "decision-actionable" uncertainty. Samples of the same ROI influence functional can also estimate sign probabilities, credible intervals, joint covariance, and the variance components in Equation 9. Moreover, Equation 7 retains the deterministic ablation effect as its mean, so improvements in mean-ranking faithfulness do not establish an advantage of the uncertainty calculation.

  2. The evidence for quantitative regional confidence is less consistent than the evidence for uncertainty ordering. Equation 8 gives a probability under the fitted parameter approximation, with the extractor fixed in the main experiments. Appendix B.3 evaluates agreement with independently retrained models, which supplies a different empirical reference. Nominal 90% intervals cover 70.8% of retrained CN-MCI effects in Table 18 and 83.1-87.1% of HMC influence draws in Table 5. Table 13 reports more favorable CN-MCI retraining coverage of 0.872 in a separate evaluation and also includes sign-frequency comparisons. The conditions supporting calibrated confidence therefore need to be made clearer.

Questions for authors
  1. Which claimed advantage over posterior sampling concerns reliability itself, and which concerns computational cost?

  2. Under what conditions should the reported confidence values and variance decomposition support quantitative reliability judgments?

3 comments

We thank Reviewer S1K5 for the careful and technically precise review. We appreciate the reviewer's recognition that the attribution target and the conditions for exact inference are stated explicitly.

W1 + Q1 — advantage over posterior sampling

Reply: We agree that sampling-based approaches (repeated stochastic passes, retrained models, or posterior draws of the same functional) can estimate the same quantities, including sign probabilities, intervals, covariance and the decomposition in Eq. 9. The difference is how many samples that takes. Monte Carlo error shrinks only as 1/S1/\sqrt{S}: a variance estimated from SS Gaussian draws has relative standard error 2/(S−1)\sqrt{2/(S-1)}, about 20% at S=50S = 50 and still 4.5% at S=1000S = 1000. Posterior-sampling practice accordingly asks for an effective sample size of at least 400 before interval estimates are trusted (Vehtari et al., 2021), and sampled explanations have been shown to change between hundreds and thousands of perturbations (Slack et al., 2021). CFIP gives the influence posterior in closed form, exact in the last-layer case (Theorem 1), so none of this sampling is needed. One advantage over sampling is therefore computational cost and the absence of Monte Carlo error. Also, what CFIP adds about the uncertainty is knowing which parts of the ranking to trust: CFIP’s variances track retraining variability far better than MC-dropout or BootstrapSHAP (correlation 0.910 vs 0.537 and 0.179, Table 13), and ROIs that enter the top ten only intermittently across retrainings carry 1.24 times the uncertainty of those always selected (Sec. 5.6).

W2 + Q2 — when the confidence values support quantitative judgments

Reply: We agree that ordering and calibrated level are different claims and will separate them in the text. Ordering is supported as reported. The calibrated level is a bit compromised for harder tasks like CN-MCI because the signal is weaker, so different training runs settle on different solutions: ranking stability across retrains drops from 0.652 on CN–AD to 0.541 on CN–MCI (Table 12). A posterior built from one trained model does not include this between-run variation, so its intervals are narrower than the spread of retrained models. Against HMC, which uses the same trained model, the gap is small: Laplace keeps the variance ordering (ρ=1.000\rho = 1.000) and its variances are 92–95% of HMC’s (Table 5). We have therefore introduced a calibration step: the uncertainty hyperparameter is selected on a held-out set so that the 90% coverage target is met, and on held-out subjects this holds across all three ADNI tasks (table). With this step the intervals meet their stated level on all three tasks.

Retraining Cov@90CN–ADAD–MCICN–MCI
Diagonal, nominal λ\lambda0.931 ± 0.0820.893 ± 0.1940.708 ± 0.171
Diagonal, coverage-calibrated0.903 ± 0.0300.908 ± 0.0240.903 ± 0.077
Full GGN, nominal λ\lambda0.924 ± 0.0510.912 ± 0.1450.793 ± 0.131
Full GGN, coverage-calibrated0.914 ± 0.0380.919 ± 0.0210.921 ± 0.059

References: Vehtari et al., 2021 (https://arxiv.org/abs/1903.08008); Slack et al., 2021 (https://arxiv.org/abs/2008.05030).

Thank you for the detailed response and additional experiments. My concerns are addressed and I increased my score accordingly

Dear Reviewer S1K5, we thank you for upgrading your score! Thank you for your engagement and your participation in the peer review process, and we deeply appreciate you helping us improve our manuscript!

Reviewer_N3wdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a probabilistic framework for neuroimaging model interpretation, expect to provide more reliable analysis as interpretation always impacted by the limited sample sizes, heterogeneous cohorts, and training uncertainty.

The method, which is called CFIP, defines regional influence as the model-output change induced by intervening on an ROI, uses a Laplace approximation. It turns ROI interventions and Laplace-linearized uncertainty into a closed-form posterior over regional influences. The mehtod is evaluated on Alzheimer’s disease and aging settings across multi-cohort neuroimaging benchmarks, including both the structural MRI-derived regional features and fMRI connectivity representations.

Critical strengths
  1. Moving from point attributions to an analytic distribution over regional effects is well-motivated for safety-critical neuroimaging.

  2. The efficency of the method is quite impressive. Unlike sampling-based uncertainty estimators (e.g., Monte Carlo dropout, Deep Ensembles, or BootstrapSHAP), CFIP avoids repetitive backpropagation or sampling at inference time. Once the precision matrix is computed, generating an influence posterior takes extremly low time per subject.

The paper also provide the code with detailed instruction to better reproduce the result.

Critical weaknesses
  1. The method is not on the imaging data, instead it is on the measurement from the third party tool. This means we are assuming that the tools are 100% realibe. It would be more useful that if we can build a end-to-end model directly on the imaging data, instead of the tabular data.

  2. Like above, as the data is on the measurement, and all main experiments are conducted using a one-hidden-layer MLP. This is highly suggest that we donot even need a network to do this.

  3. The improvement compare to the current best seems not significant.

Questions for authors
  1. What rule sets the prior precision λ so widths can be trusted? A coverage-based selection rule that brings all three ADNI tasks to ≈0.90 would remove most of the main weakness and move your draft from 5 to 6 or higher.

  2. Why is CN–MCI coverage 0.872 in Table 13 but 0.708 in Table 18, is this caused by different estimand, folds, or interval construction? If they measure the same thing, one should be wrong.

  3. Can we get more detail about how exactly was HMC run, for example, parameter block, prior, chains, draws, convergence?

  4. can we run a KFAC or full-GGN Laplace on the same endpoints: is the miscalibration the Laplace idea, or just the diagonal shortcut?

2 comments

We thank Reviewer N3wd for the careful reading and the constructive questions. We appreciate the reviewer’s recognition that moving from point attributions to an analytic distribution over regional effects is well motivated for safety-critical neuroimaging.

W1 + W2 — ROI-level features from tool, model class

Reply: CFIP targets ROI-level interpretation by design. Analysing tool-derived regional features is standard practice in neuroimaging (Fischl, 2012; Cuingnet et al., 2011; Rathore et al., 2017), regions are the unit in which findings are reported. Larger networks are also not needed for these tasks: a reproducible ADNI benchmark showed no advantage for image-level CNNs over an SVM (Wen et al., 2020), and UK Biobank benchmarking gave deep models no edge over linear ones on brain scans (Schulz et al., 2020). CFIP nevertheless applies across model sizes: logistic regression is its exact case (Theorem 1; Table 15, coverage 0.901), the same theorem holds for the last layer of any deeper network, including a CNN.

Q1 — coverage-based selection rule for λ\lambda

Reply: We appreciate the reviewer's idea of coverage-based selection for λ\lambda that can bring all bring all three ADNI tasks to 0.9. The under-coverage appears on harder tasks such as CN–MCI because the signal is weaker, so different training runs settle on different solutions: ranking stability across retrains drops from 0.652 on CN–AD to 0.541 on CN–MCI (Table 12). A posterior built from one trained model does not include this between-run variation, so its intervals are narrower than the spread of retrained models. Against HMC, which uses the same trained model, the gap is small: Laplace keeps the variance ordering (ρ=1.000\rho = 1.000) and its variances are 92–95% of HMC’s (Table 5). Following the suggestion, we now use a held-out calibration set to choose λ\lambda: among a wide range of candidate values, we keep the one whose intervals reach 0.90 coverage on the calibration subjects with the narrowest width.

Retraining Cov@90CN–ADAD–MCICN–MCI
Coverage-selected λ\lambda, full GGN0.914 ± 0.0380.919 ± 0.0210.921 ± 0.059

Q2 — Table 13 versus Table 18 mismatch

Reply: Both tables measure the same quantity, the retraining coverage; the estimand and the interval construction are identical. The difference is the hyperparameter setting. Table 13 belongs to the multi-cohort benchmark, where CFIP and the other baselines in that table were all run under one shared setting, while Tables 3 and 18 use the setting of the main ADNI analysis. We apologize for the mismatch and for not stating it clearly. In the camera-ready version we will use the same hyperparameter setting for both tables.

Q3 — HMC configuration

Reply: We ran HMC over the 65 output-layer parameters (64 weights and the bias), with the trained feature extractor fixed. The target uses the same fitting subjects, class-weighted cross-entropy and prior as the matched Laplace approximation. For each of the three ADNI tasks and five folds we ran four independent chains with 500 warm-up and 2,000 retained draws each (8,000 draws per task and fold).

Q4 — diagonal versus full GGN

Reply: We ran the full GGN on the same retraining-coverage endpoint. It generally improves coverage, so part of the miscalibration is the diagonal shortcut. It still does not bring all tasks to 0.90, so the rest is not a curvature effect. The coverage-based calibration from Q1 closes the remaining gap for both curvatures.

Retraining Cov@90CN–ADAD–MCICN–MCI
Diagonal, nominal λ\lambda0.931 ± 0.0820.893 ± 0.1940.708 ± 0.171
Diagonal, coverage-calibrated λ\lambda0.903 ± 0.0300.908 ± 0.0240.903 ± 0.077
Full GGN, nominal λ\lambda0.924 ± 0.0510.912 ± 0.1450.793 ± 0.131
Full GGN, coverage-calibrated λ\lambda0.914 ± 0.0380.919 ± 0.0210.921 ± 0.059

References: Wen et al., 2020 (https://arxiv.org/abs/1904.07773v6); Schulz et al., 2020 (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7447816/).

Dear Reviewer N3wd,

We hope we’ve addressed all your concerns. Since the discussion period ends in <6 hours, please let us know if you need any further clarification. Thank you for your engagement and your participation in the peer review process, and we deeply appreciate you helping us improve our manuscript!

Thank you, The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.