cs.CVOct 21, 2026
SaveOpenMAP-HIE: Pathology Informed Score-based Diffusion Model for Unsupervised Neonatal Hypoxic-Ischemic Encephalopathy Lesion Segmentation
Organizations: Johns Hopkins University · Johns Hopkins University, University of Alberta · Johns Hopkins University School of Medicine
Abstract
Accurate localization of neonatal hypoxic--ischemic encephalopathy (HIE) lesions on apparent diffusion coefficient (ADC) images is clinically important but is limited by scarce expert annotations and substantial variation in lesion appearance. We propose OpenMAP-HIE, a pathology-informed score-based diffusion model for unsupervised HIE lesion localization without voxel-wise lesion supervision. A denoising diffusion model learns the healthy neonatal ADC distribution, while a ControlNet-based conditional branch preserves patient-specific anatomy. At inference, anomaly evidence is derived directly from differences between conditional and unconditioned noise predictions at independently evaluated diffusion noise levels, avoiding dependence on a final pseudo-healthy reconstruction or sequential reverse-diffusion chain. Based on the characteristic ADC reduction associated with acute HIE, only discrepancies consistent with the expected pathological direction are retained and aggregated across noise levels using a geometric mean. Spatial mean filtering further suppresses isolated responses. The framework was evaluated on the BONBID-HIE dataset. The ADC-only method achieved a Dice score of , average precision of , Hausdorff distance of mm, mean average surface distance of mm, and normalized surface distance of , providing the best overlap and surface-distance performance among the evaluated unsupervised methods and Dice performance comparable to supervised ADC-only models. Optional population-derived ADC -map refinement further increased Dice to . These findings support pathology-informed score discrepancies as a useful signal for neonatal HIE lesion localization, while highlighting remaining challenges in detecting small and low-contrast lesions.
Figures & tables
Figure 1: Architecture of the anatomy-conditioned diffusion model. The healthy-prior U-Net is pretrained in stage 1 and frozen during stage 2. A trainable ControlNet encoder, initialized from the U-Net encoder, incorporates the spatial condition alongside the noisy input and timestep . Multi-scale ControlNet features are passed through zero-initialized convolutions and added to the U-Net skip features, enabling subject-specific conditioning of the noise prediction. Dashed gray lines denote U-Net skip connections, and orange arrows indicate ControlNet residual connections.
Figure 2: Illustration of inference using pathology-informed noise-prediction discrepancy aggregation for (a) a normal slice and (b) a slice containing HIE lesions. The observed ADC image is independently perturbed at multiple noise levels. At each level, the patient-conditioned (green dashed line) and healthy-prior (red dashed line) branches evaluate the same noisy input, and their predicted noise difference is retained only in the direction consistent with the reduced-ADC prior. Normal tissue is expected to exhibit relatively small discrepancies, whereas lesion regions may produce larger, directionally consistent responses across noise levels. The resulting discrepancy maps are aggregated using the geometric mean, , to obtain the anomaly map. Blue points denote noisy inputs, green and orange markers denote patient-conditioned and healthy-prior predictions, respectively, and red brackets indicate directional discrepancies. The curves provide a sequential forward-diffusion trajectory. The lesion ground truth annotation in (b) is shown for reference only.
Figure 3: Qualitative comparison of unsupervised anomaly detection methods in subjects with (a) small ( ), (b) medium ( ), and (c) large ( ) lesion burdens, defined by the subject-level lesion percentage. Columns show the input ADC image, expert lesion annotation, and anomaly maps from selected top-performing baselines, including DAE ( Kascenas et al., 2022 ) , ANDi ( Frotscher et al., 2026 ) , MAD-AD ( Beizaee et al., 2025b ) , where the originally proposed smoothing methods from each paper have been applied. Our proposed method before and after mean filtering (MF; ) is highlighted in green box. Red overlays in the ground truth (GT) column indicate annotated lesions, and cyan contours delineate their boundaries on the anomaly maps. Brighter colors indicate higher anomaly scores.
Figure 4: Qualitative comparison of reconstruction-based anomaly detection and the proposed discrepancy-based approach. (a) Input ADC image and expert lesion annotation. (b–d) Reconstructed images and corresponding anomaly maps obtained using DDPM and DDPM with ControlNet at CFG scales of 0.8 and 7.5. (e) Anomaly maps produced by the proposed method before and after mean filtering (MF, ). (f) The corresponding population-derived Z-map for the input and the Z-map guided post-processing lesion refinement. Cyan contours delineate the expert-annotated lesion boundaries. All methods’ results have been applied with HIE pathological informed sign direction.
| Method | Dice | AP | HD (mm) | MASD (mm) | NSD |
| f-AnoGAN ( Schlegl et al., 2019 ) | Unstable training and unable to achieve convergence. | ||||
| DAE ( Kascenas et al., 2022 ) | |||||
| AnoDDPM ( Wyatt et al., 2022 ) | Unstable training with simplex noise. | ||||
| AutoDDPM ( Bercea et al., 2023 ) | |||||
| pDDPM ( Behrendt et al., 2024 ) | |||||
| THOR ( Bercea et al., 2024 ) | |||||
Table 1: Quantitative comparison of unsupervised anomaly detection methods for neonatal HIE lesion localization. Values are reported as mean standard deviation where available. The best and second-best results are shown in bold and underlined , respectively. For each model’s result, the originally proposed smoothing methods have been applied. For Dice, † and ‡ indicate significantly lower scores than Ours and Ours with mean filtering ( ), respectively (under two-sided paired Wilcoxon signed-rank tests, all , ).
| Lesion burden | Post-processing | Dice | AP | HD (mm) | MASD (mm) | NSD |
| Small ( ), | ||||||
| MF5 | ||||||
| Medium ( – ), | ||||||
| MF5 | ||||||
| Large ( ), | 0.541 | |||||
| MF5 |
Table 2: Performance of our proposed model stratified by subject-level lesion burden. Subjects were divided into small ( ), medium ( – ), and large ( ) lesion-burden groups according to the percentage of voxels labeled as lesions in each subject. MF5 denotes mean filtering with a kernel size of (rows shaded in gray). Results are reported as mean standard deviation where available, and the best result within each lesion-burden group is highlighted in bold.
| Model | Input or refinement | Dice | AP | HD (mm) | MASD (mm) | NSD |
| Supervised segmentation methods | ||||||
| Swin-UNETR ( Almansour et al., 2025 ) | ADC only | |||||
| ADC + -map | ||||||
| ADC + -map + | ||||||
| UNet++ ( Rao et al., 2025 ) | ADC only | |||||
| ADC + -map | ||||||
Table 3: Effect of incorporating population-derived ADC -map information. ADC-only models serve as the reference configurations (rows shaded in gray). For the supervised models, the -map and a thresholded binary -mask ( ) are provided as additional input channels ( Almansour et al., 2025 ; Bao et al., 2025 ; Rao et al., 2025 ) . For the population-derived statistical baseline, is used as the continuous anomaly score for AP, whereas is used to obtain the binary mask for Dice and surface-distance metrics. Our unsupervised method uses the -map only during post processing. AP is not reported for this post-processing result because it directly produces a binary lesion mask. The best result within each model is shown in bold . For Dice, † indicates a significantly lower score than our proposed method with -map post-processing (under two-sided paired Wilcoxon signed-rank test at significance level , ).
| Configuration | Dice | AP | HD (mm) | MASD (mm) | NSD |
| DDPM | |||||
| ControlNet (healthy-restoration, CFG=0.8) | |||||
| ControlNet (lesion-emphasizing, CFG=7.5) | |||||
| Multi noise-level discrepancy accumulation | |||||
| Mean filtering ( ) |
Table 4: Ablation study of the proposed framework. All compared models applied the HIE pathology informed directional information. Results are reported as mean standard deviation where applicable. The best and second-best results are shown in bold and underlined , respectively.
| Kernel size | Dice | AP | HD (mm) | MASD (mm) | NSD |
| 0 | |||||
| 3 | |||||
| 5 | |||||
| 7 | |||||
| 8 | |||||
| 10 |
Table 5: Ablation study of the mean-filter kernel size applied to the direction-aware anomaly map. Kernel size indicates that no mean filtering was applied. The gray row denotes the kernel size used in the main experiments. Results are reported as mean standard deviation where applicable, and the best value for each metric is highlighted in bold .
| Score representation | Evaluation | Dice | AP | HD (mm) | MASD (mm) | NSD |
| Independent | ||||||
| MF5 | Independent | |||||
| Reverse chain | ||||||
| MF5 | Reverse chain | |||||
| using forward noise | Independent |
Table 6: Ablation of the anomaly-score representation compared with sequential reverse propagation and discrepancy to forward process noise. “Independent” indicates that each noise level was sampled directly from the original patient ADC image without propagating a reverse-diffusion chain, and “Reverse chain” evaluates successive states along a sequential reverse-diffusion trajectory (Eq. ( 43 )). The final row replaces the learned unconditional noise prediction with the Gaussian noise sampled during forward corruption (Eq. ( 44 )). The biological directional constraint was enabled for all configurations. Rows shaded in gray use a mean filtering with kernel size of (MF5). Values are reported as mean standard deviation where available, and the best result for each metric is highlighted in bold .
| Statistic | Lesion Mean (95% CI) | Background Mean (95% CI) |
| Positive discrepancy probability, | ||
| Mean positive discrepancy, | ||
| Mean negative discrepancy, | ||
| Mean discrepancy magnitude, |
Table 7: Directional characteristics of the discrepancy across the selected noise levels. The discrepancy is defined as . A positive value indicates that the healthy-prior response predicts a higher ADC value than the patient-conditioned response, consistent with the expected direction of reduced ADC in acute HIE injury. Values are reported as the timestep-wise mean and 95% confidence interval across the selected noise levels ( ).
| HIE Pathology prior directional sign | Post-processing | Dice | AP | HD (mm) | MASD (mm) | NSD |
| No | ||||||
| No | MF5 | |||||
| Yes | ||||||
| Yes | MF5 |
Table 8: Ablation of the pathology-informed directional sign constraint and mean-filter post-processing. MF5 denotes mean filtering with a kernel size of 5. Rows using MF5 are shaded, and the best result for each metric is shown in bold .
| Model | Dice | AP | HD (mm) | MASD (mm) | NSD |
| DAE ( Kascenas et al., 2022 ) | |||||
| AutoDDPM ( Bercea et al., 2023 ) | |||||
| pDDPM ( Behrendt et al., 2024 ) | |||||
| MAD-AD ( Beizaee et al., 2025b ) |
Table 9: Quantitative comparison of unsupervised anomaly-detection baselines with introducing HIE pathology directional prior information at final reconstruction. Values are reported as mean standard deviation where available. The best result for each metric is highlighted in bold . For each model’s result, the originally proposed smoothing methods have been applied. For Dice, † and ‡ indicate significantly lower scores than Ours and Ours with mean filtering ( ), respectively (under two-sided paired Wilcoxon signed-rank tests, all , ).
More in this topic
Medical Image Segmentation
Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.
See all 788 papers