OpenMAP-HIE: Pathology Informed Score-based Diffusion Model for Unsupervised Neonatal Hypoxic-Ischemic Encephalopathy Lesion Segmentation

1. Introduction

• Small, multifocal, diffuse, heterogeneous lesions and costly expert delineation make automated HIE localization challenging.Scanner and acquisition variation, developmental anatomy, and ambiguous lesion boundaries further complicate distinguishing pathology from normal variability.Source: §1 · pp. 1–2
• OpenMAP-HIE learns a healthy neonatal ADC prior and uses a ControlNet branch conditioned on the observed image to capture patient-specific anatomy without lesion masks as voxel-wise optimization targets.The healthy-prior and subject-conditioned responses are compared to derive lesion evidence.Source: §1 · pp. 2–3
• Both branches evaluate the same noisy input at independently selected noise levels, deriving anomaly evidence without a final pseudo-healthy reconstruction or sequential reverse-diffusion chain.The independently evaluated discrepancies are aggregated across noise levels.Source: §1 · pp. 2–3
• The method retains discrepancies consistent with acute HIE's characteristic ADC reduction and suppresses those in the opposite direction; population-derived ADC Z-maps are optional.The core framework requires only the subject's ADC image at inference.Source: §1 · pp. 2–3
• The ADC-only method improved lesion overlap and surface localization over evaluated unsupervised baselines, while directional filtering provided substantial localization gains.Population-derived Z-map refinement further improved overlap and boundary accuracy when population-level normative information was available.Source: §1 · p. 3

2.1. Supervised Anomaly Detection

• Supervised HIE segmentation learns lesion appearance from voxel-wise annotations, whereas this framework learns a healthy-image prior without lesion masks as optimization targets.Supervised systems include U-Net variants, SegResNet, Swin-UNETR, ensembles, and imbalance-oriented losses.Source: §2.1 · p. 3

2.2. Unsupervised Anomaly Detection

• Reconstruction-based methods rely on larger errors in pathological regions, but imperfect healthy reconstruction or reproduction of pathology means fidelity alone does not ensure pathological specificity.These competing effects can undermine residual-based anomaly evidence.Source: §2.2 · p. 4

2.3. Diffusion Model and Score Method

• Diffusion anomaly methods include reconstruction and inpainting approaches; ANDi instead aggregates discrepancies across noise levels, while OpenMAP-HIE compares subject-conditioned and healthy-prior score estimates.Unlike ANDi's model-prediction versus analytical-posterior comparison, OpenMAP-HIE also applies an HIE-specific directional filter.Source: §2.3 · p. 4

2.4. HIE Pathology Directional Prior

• Acute HIE typically lowers ADC in injured tissue, so OpenMAP-HIE retains only positive differences corresponding to higher healthy-reference than subject-conditioned reverse-mean estimates.This suppresses large denoising discrepancies inconsistent with the characteristic ADC reduction.Source: §2.4 · p. 5

3.1. Diffusion Model

• The healthy-prior denoiser is pretrained on healthy neonatal ADC slices, with its noise prediction parameterizing the score of the perturbed image distribution.The model progressively perturbs healthy slices and is trained to predict the noise removed by the reverse process.Source: §3.1 · p. 5

3.2. Spatial Conditioning through ControlNet

• ControlNet conditions denoising on the observed ADC image to preserve subject-level anatomy, while the pretrained healthy-prior U-Net remains frozen during ControlNet training.Conditioning inputs are randomly dropped during training so the model can provide both null-conditioned and subject-conditioned predictions.Source: §3.2 · p. 5

3.3. Classifier-Free Guidance and Interpretation of the Conditional–Unconditional Discrepancy

• The conditional–unconditional discrepancy measures how introducing the observed image as spatial conditioning changes the denoising response.A larger discrepancy may reflect pathology, but may also arise from normal anatomical variation, registration differences, acquisition-related intensity variation, or artifacts.Source: §3.3 · p. 6
• Changing the classifier-free guidance scale only scales the discrepancy at a fixed noise level; the discrepancy itself is independent of that scale.The scale can still affect a complete reverse-diffusion trajectory and reconstruction-based anomaly maps.Source: §3.3 · p. 6

3.4. Pathology-Informed Directional Discrepancy and Multi Noise-Level Accumulation

• The directional prior retains discrepancies consistent with restoring lower subject-conditioned ADC toward higher healthy-prior ADC, matching the expected ADC reduction in acute HIE.The noise-prediction discrepancy is used directly, without timestep coefficient weighting or selecting a classifier-free guidance scale.Source: §3.4 · pp. 6–7
• Each evaluated noise level is constructed independently from the same observed image, and geometric-mean aggregation favors directional responses that remain elevated across levels.The positive-part operation excludes opposite-direction discrepancies before accumulation; geometric aggregation reduces the influence of isolated large responses.Source: §3.4 · p. 7
• The directional anomaly evidence does not establish lesion specificity, because normal anatomical variation, partial-volume effects, registration differences, intensity variation, and artifacts can also produce positive discrepancies.The score represents persistent deviation directionally consistent with acute HIE-related ADC reduction, not a calibrated lesion probability.Source: §3.4 · p. 7

3.5. Spatial Smoothing of the Anomaly Map

• Mean filtering reduces isolated high-valued responses while leaving spatially coherent regions relatively elevated; the final framework uses a validation-selected kernel size of k=5.An excessively large kernel may dilute focal lesions and spread elevated scores into adjacent healthy tissue, increasing false positives.Source: §3.5 · p. 8

3.6. Population Z-Map Guided Post-processing Lesion Refinement

• Population Z-map refinement identifies a high-confidence low-ADC core, then considers high-anomaly voxels with negative population deviation for diffusion-supported component recovery.The high-anomaly threshold is the subject-specific 99th percentile; candidate components must exceed 10 voxels and be adjacent to the cleaned Z-map core.Source: §3.6 · p. 8
• The fusion uses the population Z-map to localize spatially coherent low-ADC tissue and diffusion discrepancy to recover nearby lesion components, without manual lesion annotations at inference.The diffusion score supplies complementary subject-specific evidence rather than independently adding distant high-anomaly regions.Source: §3.6 · p. 8

3.7. Dataset and Preprocessing

• BONBID-HIE was divided into 85 training, 4 validation, and 44 test subjects, with all 133 subjects included and no overlap across partitions.Evaluation used complete 3D volumes from the 44 test subjects.Source: §3.7 · pp. 8–9
• The diffusion model was trained on 6,042 axial slices without annotated lesion voxels; lesion masks selected training slices but were not model inputs or optimization targets.Population Z-maps were introduced only during post-processing, not to train the diffusion model.Source: §3.7 · p. 9

3.8. Network Architecture and Implementation Details

• The pretrained U-Net was frozen during the second training stage, while zero-initialized ControlNet residual projections learned subject-specific spatial information.The zero initialization preserved the pretrained U-Net’s healthy denoising behavior at the start of the second stage.Source: §3.8 · p. 9
• Inference accumulates discrepancies over timesteps 75–199, selected to weaken lesion signal while retaining enough anatomy for spatially meaningful comparison.Lower noise levels provide insufficient perturbation, while substantially higher levels remove subject-specific anatomical information.Source: §3.8 · p. 9

3.9. Evaluation Metrics

• Metrics were computed over each subject’s complete 3D volume and reported as the mean ± standard deviation across test subjects.The evaluation covers voxel ranking, volumetric overlap, and surface distance.Source: §3.9 · p. 9
• Dice measures volumetric overlap, ranges from 0 to 1, and is higher when overlap is greater.Source: §3.9 · pp. 9–10
• The reported ceiling Dice uses one threshold maximizing mean subject-wise test-set Dice, selected with test annotations; the same threshold is used for surface metrics.Surface metrics receive no further threshold optimization, so these binary metrics use an oracle threshold protocol.Source: §3.9 · p. 10
• Average precision summarizes the precision–recall curve over continuous anomaly scores and is independent of the selected threshold.It is particularly informative because lesion voxels make up a small proportion of the evaluated brain volume.Source: §3.9 · p. 10
• Hausdorff distance measures maximum surface disagreement, while mean average surface distance measures mean bidirectional distance between lesion surfaces.Lower values indicate smaller maximum boundary error for Hausdorff distance and closer average boundary agreement for mean average surface distance.Source: §3.9 · p. 10
• Normalized surface distance is the proportion of surface voxels within a 2.0 mm tolerance of the corresponding surface; higher values indicate better agreement.Surface distances are expressed in millimeters, with coordinates accounting for voxel spacing.Source: §3.9 · p. 10

3.10. Statistical Analysis

• Two-sided paired Wilcoxon signed-rank tests used each test subject as the paired unit, with Dice as the primary outcome and significance set at α = 0.05.Mean paired Dice differences accompany p-values; ceiling-Dice comparisons use the test-set oracle threshold protocol.Source: §3.10 · p. 11

3.11. Alternative Discrepancy Formulations for Ablation

• The reverse-chain ablation tests whether a dependent reverse trajectory is preferable to independently evaluating matched corruption levels of the observed image.The final method evaluates each selected noise level independently.Source: §3.11 · p. 11
• The forward-noise-reference ablation tests whether conditional noise-prediction error alone can provide anomaly evidence without a learned unconditional healthy reference.Source: §3.11 · p. 11

4.1. Unsupervised Benchmark Results

• With k=5, Dice was 0.283, AP was 0.505, HD was 25.45 mm, MASD was 9.40 mm, and NSD was 0.355, with the best Dice, HD, MASD, and NSD among successfully trained methods.Relative to MAD-AD, Dice improved by 0.117, HD fell by 1.99 mm, MASD by 1.93 mm, and NSD rose by 0.142; DAE retained the highest AP at 0.665.Source: §4.1 · p. 11, Table 1 · p. 12
• MF5 increased Dice from 0.080 to 0.141 for small-lesion subjects and from 0.287 to 0.377 for medium-lesion subjects, while large-lesion Dice was 0.603 versus 0.600.The small-lesion example remained difficult: the lesion was detected but less noticeable amid surrounding false-positive responses.Source: §4.1 · pp. 11, 13

4.2. Comparison with Supervised ADC-Only Models

• In the ADC-only comparison, OpenMAP-HIE’s Dice of 0.232 was comparable to UNet++ at 0.237, but its AP of 0.374 and HD of 29.29 mm lagged both supervised models.Swin-UNETR had the highest ADC-only Dice and NSD, at 0.311 and 0.406, respectively.Source: §4.2 · p. 13

4.3. Effect of Combining Population-Derived ADC 𝑍-Map Information

• Adding Z-map-guided post-processing increased Dice from 0.232 to 0.602 and NSD from 0.284 to 0.667; Dice was significantly higher than the ADC-only prediction.The refinement uses the Z-map only after diffusion prediction to identify a low-ADC core and retain nearby diffusion-supported components.Source: §4.3 · p. 13, Table 3 · p. 14

4.4. Ablation Analysis of Framework Components

• Multi-noise-level discrepancy accumulation improved Dice from 0.167 to 0.232 and reduced HD from 32.03 to 29.29 mm, whereas single-reconstruction anomaly maps showed limited and inconsistent lesion sensitivity.The discrepancy is computed directly from conditional and unconditional noise predictions, so its magnitude is independent of CFG scale; MF5 produced the best performance across metrics.Source: §4.4 · pp. 14–15, Table 4 · p. 15

4.5. Ablation Analysis of Mean Filtering Kernel Size

• Among tested kernels, k=5 had the highest AP (0.505), lowest HD (25.45 mm) and MASD (9.40 mm), and highest NSD (0.355); larger kernels could dilute focal responses.k=7 had slightly higher Dice than k=5 (0.287 versus 0.283), but inferior AP and surface-distance metrics.Source: §4.5 · p. 15, Table 5 · p. 15

4.6. Ablation Analysis of Discrepancy

• Independent evaluation outperformed reverse-chain accumulation in Dice (0.232 versus 0.070), while replacing the learned unconditional prediction with sampled forward noise yielded Dice 0.045.The authors attribute the reverse-chain disadvantage to propagated errors and attenuation of pathological information; the unconditional branch supplies a learned healthy-prior reference.Source: §4.6 · p. 16, Table 6 · p. 17

4.7. Verification of the HIE Pathology Directional Prior

• With MF5, the pathology-direction constraint increased Dice from 0.209 to 0.283 and improved all evaluated metrics; lesion voxels showed positive discrepancy in 88.6% of evaluated voxel–timestep pairs.OpenMAP-HIE with MF5 also exceeded each direction-filtered baseline in Dice, with p<0.001 for all comparisons.Source: §4.7 · pp. 16, 18, Table 8 · p. 17

5.1. Overall Detection Performance

• ADC-only OpenMAP-HIE showed strong Dice and surface localization, but lower AP indicates imperfectly calibrated scores with lesion and background values overlapping.The authors describe healthy-prior modeling as a complementary route to quantitative HIE assessment, particularly when annotations do not represent lesion diversity.Source: §5.1 · p. 18

5.2. Intermediate Denoising Discrepancy versus Reconstruction Difference

• Intermediate conditional–unconditional discrepancies retain subject conditioning, avoid reverse-chain error propagation, and emphasize evidence persistent across noise levels.The authors contrast this with terminal reconstruction differences, which can mix lesion restoration with stochastic variation and unintended changes to healthy anatomy.Source: §5.2 · p. 19

5.3. Pathology-Informed Directionality and Interpretability

• Retaining discrepancies consistent with reduced ADC makes the score pathology-informed; annotated lesions showed larger positive discrepancies, and the sign constraint improved localization over unsigned scoring.Applying the same low-ADC rule after reconstruction also improved conventional outputs, but less than integrating it into multi-noise discrepancy scoring.Source: §5.3 · pp. 19–20

5.4. Utility and Limitations of Population 𝑍-Score Refinement

• Population-derived ADC Z-maps can improve localization, but depend on appropriate reference statistics and accurate registration, limiting transfer across populations and acquisition settings.The ADC-only model avoids voxel-wise population statistics and precise nonlinear normalization, though its localization performance is lower than the Z-map-assisted configuration.Source: §5.4 · p. 20

5.5. Limitations and Future Work

• Training used lesion masks to exclude slices, evaluation used one cohort and fixed preprocessing, and 2D inference omits through-plane context.The authors call for clinically healthy training data, multicenter validation, and 2.5D or 3D architectures.Source: §5.5 · p. 20
• The reduced-ADC assumption may miss injury after pseudo-normalization, while hard clipping removes weak or near-zero discrepancies and may reduce sensitivity to small or low-contrast lesions.The continuous anomaly score is also not a calibrated lesion probability and can vary with noise realizations and selected noise levels.Source: §5.5 · pp. 20–21

• OpenMAP-HIE localizes neonatal HIE abnormalities on ADC images without lesion annotations as voxel-wise training targets.Source: §6 · p. 21
• The ADC-only method improved lesion overlap and surface-distance performance over evaluated unsupervised baselines and achieved Dice performance comparable to supervised ADC-only models.Source: §6 · p. 21
• Population-derived ADC Z-map refinement provided further improvements in lesion localization.Source: §6 · p. 21

• Pengyu Kan is credited with software, visualization, investigation, and original drafting; Craig Jones with methodology and formal analysis; and Kenichi Oishi with funding, project administration, and supervision.Source: CRediT authorship contribution statement · p. 21

• Kenichi Oishi is a consultant for Corporate-M and AnatomyWorks, with the arrangements managed under Johns Hopkins University conflict-of-interest policies.Source: Declaration of competing interest · p. 21