Source-linked AI summary
DuoPIC: Uncertainty-Aware Voxel-Level Digital Twins of Cerebral Perfusion with Transport PDEs in Acute Stroke
Anh Duong Tran, Quang-Anh N. D., Toan Sy Pham, Trang H. Hoang, Minh N. Do, Duy Ton Mai, Huan X. Nguyen
TL;DR
Acute-stroke perfusion decisions rely on patient-specific CBF, CBV and MTT, while existing approaches have limitations in robustness, physiological explanation and uncertainty. DuoPIC combines patient-specific haemodynamics with tracer transport and a shared neural closure; its voxel maps achieve state-of-the-art lesion detection across ISLES’17 MR and ISLES’24 CT.
Problem
Reliable patient-specific CBF, CBV and MTT estimates matter for treatment decisions, but current approaches have limitations in robustness, physiological interpretation, or uncertainty quantification.
Source: §1 · p. 1Method
DuoPIC couples patient-specific haemodynamics with tracer transport and uses a shared neural closure to model regional transit-time heterogeneity while preserving physiological quantities.
Source: §1 · p. 1, §2 · p. 1, §2.2 · p. 3Results
Takeaways & Limitations
The framework provides reference-free, uncertainty-aware voxel estimates and better predicts unseen frames while retaining physiological CBF, CBV and MTT.
Source: §4 · p. 4, §5 · p. 4A single tube narrows the transit-time distribution, whereas ROIs mixing fast and slow tissue require a wider distribution.
Source: §2.2 · p. 3
Abstract
from arXiv · showhide
Digital twins of cerebral perfusion relate tracer dynamics to cerebral blood flow (CBF), blood volume (CBV) and mean transit time (MTT), but must be data-driven, physically consistent and uncertainty-aware. We introduce DuoPIC, a digital twin that needs no reference maps, in which a Windkessel flow model drives a tracer transport partial differential equation (PDE). A population-level neural closure divides each region into fast and slow pathways with tracer mass and MTT conserved by construction, and a tempered posterior quantifies the uncertainty of every estimate. Its voxel maps set the state of the art for lesion detection on ISLES'17 MR (AUC 0.745) and ISLES'24 CT (0.743). On held-out patients, the neural PDE reduced the compartment's reconstruction error by 19% and lifted 95% MTT interval coverage from 88% to 94%. Code: https://github.com/detinatran/duopic
1. INTRODUCTION
Reliable patient-specific perfusion estimates matter for acute stroke decisions, but existing approaches face limits in robustness, physiological interpretation, or uncertainty quantification. DuoPIC addresses these gaps with a reference-free, uncertainty-aware perfusion digital twin and a shared neural closure that preserves physiological quantities.
1. INTRODUCTION
- Perfusion imaging reveals hypoperfused but salvageable tissue, and treatment decisions depend on reliable patient-specific CBF, CBV and MTT estimates.In DEFUSE 3, perfusion imaging selected patients benefiting from thrombectomy 6–16 hours after last known well.Source: §1 · p. 1
- Clinical deconvolution is sensitive to noise and the arterial input function, while direct infarct-prediction networks provide no physiological quantities.Source: §1 · p. 1
- Mechanistic models have simplifying assumptions, while neural networks need labels and lack physical explanation; physics-informed learning lets networks learn what the physical model leaves unexplained.Source: §1 · p. 1
- DuoPIC is a reference-free, uncertainty-aware perfusion digital twin that couples patient-specific haemodynamics with transport physics from region to voxel level.Its shared neural closure learns transit-time heterogeneity while conserving tracer mass, CBF, CBV and MTT.Source: §1 · p. 1
- HemoPIC relied on reference deconvolution priors, could not capture capillary transit-time heterogeneity under its well-mixed regional assumption, and lacked uncertainty quantification.Source: §1 · p. 1
2. PROPOSED DUOPIC
DuoPIC estimates voxel-level CBF, CBV and MTT without reference maps by combining a regional flow model with neural transport and voxel fitting. Its neural closure conserves tracer mass and MTT, while posterior sampling and ensembles quantify uncertainty.
2. PROPOSED DUOPIC
- DuoPIC estimates voxel CBF, CBV and MTT without reference maps, using regional ROIs because voxel curves are too noisy for the regional model.It builds on HemoPIC’s clustering of grey- and white-matter voxels into homogeneous regions.Source: §2 · p. 1
2.1. Reference-free regional model
- The regional model uses Windkessel outflow and an arterial input with ROI-specific delay and dispersion, with scan-specific timing and absolute-unit priors.The arterial input is selected from clean, high-area bolus curves and corrected for partial volume using the venous output function.Source: §2 · p. 2, §2.1 · p. 2
2.2. Transport PDE and neural closure
- A shared neural closure splits each ROI into fast and slow transport tubes, conserving tracer mass and mean transit time while allowing transit-time spread to increase.The split is learned across patients from label-free descriptors because one curve cannot identify it alongside the physical parameters.Source: §2.2 · p. 3
2.3. Voxel-level fit
- Voxel fitting estimates each voxel’s amplitude, shape and delay around its ROI model, then combines shape-derived MTT with indicator-dilution CBV to obtain CBF.The ROI model fixes parameters that noisy voxel curves cannot identify; the batched fitting procedure handles about 2·10^5 voxels per patient in about 100 seconds on one CPU thread.Source: §2.3 · p. 3
2.4. Uncertainty quantification
- Uncertainty estimates combine posterior sampling of ROI and PDE parameters with voxel-level Laplace approximations, bootstrap networks and a pipeline ensemble.The checks use hidden frames, uncertain voxels and semi-synthetic refits.Source: §2.4 · p. 3
3. EXPERIMENTAL SETUP
The evaluation used ISLES’17 DSC perfusion data and an external ISLES’24 CT cohort, with lesion detection assessed in a subset of ISLES’17 patients. It also reports preprocessing, metrics, implementation, and a representative patient-map figure.
3. EXPERIMENTAL SETUP
- The study included 43 ISLES’17 subjects with DSC perfusion and infarct masks, plus 149 ISLES’24 CT subjects with follow-up MRI masks (145 with infarct in view).ISLES’17 lesion detection used 36 subjects whose AIF was earlier, narrower, and higher than the tissue mean.Source: §3 · pp. 3–4
- Figure 4 compares CBF, CBV, and MTT maps from the ISLES maps, HemoPIC, and DuoPIC for one ISLES’17 patient, showing the infarct in red and each map’s lesion AUC.The caption marks the best AUC in bold.Source: Figure 4 · p. 4
- All methods used non-contrast CT registration, baseline subtraction, and in-plane 3-voxel Gaussian smoothing; ISLES’17 preprocessing followed HemoPIC with four SynthSeg tissue classes.The study reported tracer-reconstruction nRMSE, lesion ROC AUC and average precision for CBF, CBV, and MTT relative to the brain median, and tissue-model AIC.Source: §3 · pp. 3–4
- Held-out patients were refitted for 1000 steps, and posterior sampling used four chains of 6000 steps after 3000 burn-in, retaining every fourth draw (1500 draws per chain).Source: §3 · p. 4
4. EXPERIMENTAL RESULTS
DuoPIC achieved the strongest reported lesion-detection results across ISLES’17 MR and ISLES’24 CT. The neural PDE also improved held-out curve prediction and brought uncertainty intervals closer to nominal coverage.
4. EXPERIMENTAL RESULTS
- On held-out patients, the neural PDE lowered brain nRMSE in all five folds and predicted hidden-frame curves better than the compartment in 65% of ROIs.The neural PDE improved the MTT and relaxation-time interval coverage toward 95%; Table 1 reports semi-synthetic parameter coverage, median 95% interval width relative to value, and ROI-median hidden-frame nMSE.Source: §4 · p. 4, Table 1 · p. 4
- The neural PDE narrowed the median 95% CBF interval from 116% of the value with HemoPIC to 72%, while network uncertainty was 0.6% of CBF versus 16.9% posterior uncertainty.Removing the most uncertain voxels raised MTT lesion AUC from 0.755 to 0.787.Source: §4 · p. 4
- With consistent units, data constrained CBF and CBV in over 90% of ROIs, while the compartment missed 22–37% of signal at the bolus peak.In normal grey matter, median CBF, CBV, and MTT were 52 ml/100 g/min, 4.2 ml/100 g, and 4.9 s, respectively.Source: §4 · p. 4
- Across 145 ISLES’24 CT patients, DuoPIC reached mean AUC 0.743, versus 0.691 for the best oSVD setting, 0.688 for HemoPIC, and 0.687 for vendor maps (p < 10−20 each).Source: §4 · p. 4
5. CONCLUSION
DuoPIC provides a reference-free, uncertainty-aware perfusion digital twin. Its neural transport PDE predicts unseen frames with physical CBF, CBV, and MTT, and tuning-free voxel fitting achieves state-of-the-art lesion AUC and AP; absolute values remain dependent on the automatic AIF, and the reference infarct reflects treatment.
5. CONCLUSION
- DuoPIC makes a perfusion digital twin reference-free and uncertainty-aware, with a neural transport PDE that better predicts unseen frames while retaining physical CBF, CBV, and MTT.Its tuning-free voxel fit achieves state-of-the-art lesion AUC and AP; absolute values still depend on the automatic AIF, and the reference infarct reflects treatment.Source: §5 · p. 4
6. COMPLIANCE WITH ETHICAL STANDARDS
The study used open-access ISLES’17 and ISLES’24 data, whose licenses required no ethical approval.
6. COMPLIANCE WITH ETHICAL STANDARDS
- The open-access ISLES’17 and ISLES’24 data licenses required no ethical approval.Source: §6 · p. 5