A deep learning pipeline for mapping in situ network-level neurovascular coupling in multi-photon fluorescence microscopy
Hypotheses The paper advances two coupled bets: an engineering hypothesis that a deep-learning pipeline combining segmentation, longitudinal registration, radius estimation, and graph-theoretic analysis can recover network-level neurovascular measurements unavailable to conventional single-vessel approaches, and a biological hypothesis that optogenetic neuronal activation drives coordinated, graph-level microvascular responses that point-caliber measurement cannot resolve.
Claims The NOVAS3D UNet/UNETR ensemble outperforms ilastik on Dice, precision, and HD95; registration nearly doubles the tracked vessel count (241 to 412 per field) and radius estimation recovers simulated ground truth at R-squared = 0.68. Applied to Thy1-ChR2 cortex, the pipeline reveals baseline within-vessel radius heterogeneity of 24 +/- 28%, dilations occurring closer to labeled neurons than constrictions (16.1 vs 21.9 micrometers), dilators biased toward the surface, a 152 +/- 65% rise in network assortativity, and a 4% median gain in capillary-network efficiency under 4.3 mW/mm-squared photostimulation, all absent in green-light and wild-type controls.
Inferences The within-vessel heterogeneity and network coordination findings jointly rule out the assumption that better single-vessel segmentation alone can substitute for graph-level analysis of functional hyperemia, reframing neurovascular coupling as an intrinsically network-level phenomenon. Scope is bounded by the single-preparation training set and the untested 2-SD responder threshold, which condition all quantitative biology claims.
▸ Summary
▸Hypotheses tested
A DL segmentation + registration + graph-analysis pipeline can deliver automated network-level neurovascular-coupling measurements across hundreds of vessels per FOV, where prior baselines (ilastik, point-caliber, single-time-point) fall short.
The DL pipeline should beat ilastik on segmentation metrics, registration + mask union should roughly double recovered vessel counts per FOV, and radius estimation should match simulated ground truth at high R² and tolerate realistic noise.
Tested by
The NOVAS3D deep learning segmentation pipeline (UNet/UNETR) achieves significantly higher Dice scores and precision/recall for volumetric vessel segmentation than the baseline ilastik classifier on the deposited two-photon microscopy test dataset.
The pipeline's boundary-detection radius estimation achieves R²=0.68 against ground-truth simulated radii across >100,000 simulations spanning 0.5×–2× dilation/constriction, and remains stable under Gaussian noise up to a standard deviation of 200 signal units (image range 0–1023 SU).
Rigid registration across time points and union of segmentation masks increases the number of identified vessel segments per FOV from 241±174 (single time point) to 412±281 (union across all time points; 507×507×250 µm, n=107 FOVs), while reducing mean squared error between acquisitions from 1306±747 to 0.008±0.003 signal units.
UNETR ensemble segmentation shows significantly better HD95 surface distance than ilastik for both vessel and neuron channels (p<0.05 Wilcoxon signed-rank), while ilastik over-segments vessels with high recall (0.89±0.19) but low precision (0.37±0.33), evaluated on nine test images (507×507×250 µm) from six held-out mice.
Under ChR2 photostimulation we should see a dilation-vs-constriction spatial gradient relative to neurons, network-wide assortativity rise, capillary efficiency change — and none of these in green-light or wild-type controls. Point measurements should miss all of them.
Tested by
At 1.1 mW/mm² 458 nm stimulation, a sample artery dilated 1.33±0.86 µm (p<1e-4) and a sample capillary dilated 0.42±0.39 µm (p<1e-4), while a sample venule showed no significant radius change (p=0.22), demonstrating vessel-type heterogeneity in optogenetic neurovascular responses.
Capillary radius varies along vessel length by 24±28% of the mean resting radius across baseline frames, such that point measurements of vessel caliber cannot accurately estimate microvessel volume changes.
Capillary dilations following 458 nm ChR2 stimulation (0.90±0.93 µm at 1.1 mW/mm²; 0.90±0.78 µm at 4.3 mW/mm²) are significantly larger than those following 552 nm control illumination (0.58±0.92 µm; p<1e-4), confirming that the vascular responses are ChR2-mediated rather than photothermal artifacts.
Capillary network efficiency shows a median 4% increase during peak optogenetic stimulation, consistent with coordinated vasodilation improving local blood flow distribution.
Constricting capillaries are located on average 37±179 µm deeper in cortex than dilating capillaries following 458 nm optogenetic stimulation at 4.3 mW/mm² (p<1e-4), with dilators tending toward the cortical surface across all stimulation conditions.
Following ChR2 optogenetic activation at 458 nm, capillary dilations occur on average 16.1±14.3 µm from the nearest labeled pyramidal neuron while constrictions occur 21.9±14.6 µm away (4.3 mW/mm²; p<1e-4), and this spatial segregation is absent under control 552 nm illumination.
Vascular network assortativity increases by 152 ± 65% at 4.3 mW/mm² optogenetic stimulation relative to control, indicating that high-degree vessels preferentially couple with high-degree vessels during neurovascular responses.
Vessel radius adjustments during optogenetic stimulation show 24 ± 28% variation relative to resting diameter, with dilations averaging 16.1 ± 14.3 µm and constrictions averaging 21.9 ± 14.6 µm relative to nearby neurons.
Wild-type C57BL/6J mice (n=4) show no statistically distinguishable capillary radius distributions following blue (458 nm) versus green (552 nm) photostimulation, confirming that vascular responses in Thy1-ChR2 mice are ChR2-specific and not attributable to photothermal or non-specific light effects.
Optogenetic neuronal activation drives coordinated, network-level vascular responses (spatially organised dilations/constrictions, neighbour-correlated capillary responses, stimulation-modulated efficiency) — point measurements of individual vessels cannot resolve this coordination.
Under ChR2 photostimulation we should see a dilation-vs-constriction spatial gradient relative to neurons, network-wide assortativity rise, capillary efficiency change — and none of these in green-light or wild-type controls. Point measurements should miss all of them.
Tested by
At 1.1 mW/mm² 458 nm stimulation, a sample artery dilated 1.33±0.86 µm (p<1e-4) and a sample capillary dilated 0.42±0.39 µm (p<1e-4), while a sample venule showed no significant radius change (p=0.22), demonstrating vessel-type heterogeneity in optogenetic neurovascular responses.
Capillary radius varies along vessel length by 24±28% of the mean resting radius across baseline frames, such that point measurements of vessel caliber cannot accurately estimate microvessel volume changes.
Capillary dilations following 458 nm ChR2 stimulation (0.90±0.93 µm at 1.1 mW/mm²; 0.90±0.78 µm at 4.3 mW/mm²) are significantly larger than those following 552 nm control illumination (0.58±0.92 µm; p<1e-4), confirming that the vascular responses are ChR2-mediated rather than photothermal artifacts.
Capillary network efficiency shows a median 4% increase during peak optogenetic stimulation, consistent with coordinated vasodilation improving local blood flow distribution.
Constricting capillaries are located on average 37±179 µm deeper in cortex than dilating capillaries following 458 nm optogenetic stimulation at 4.3 mW/mm² (p<1e-4), with dilators tending toward the cortical surface across all stimulation conditions.
Following ChR2 optogenetic activation at 458 nm, capillary dilations occur on average 16.1±14.3 µm from the nearest labeled pyramidal neuron while constrictions occur 21.9±14.6 µm away (4.3 mW/mm²; p<1e-4), and this spatial segregation is absent under control 552 nm illumination.
Vascular network assortativity increases by 152 ± 65% at 4.3 mW/mm² optogenetic stimulation relative to control, indicating that high-degree vessels preferentially couple with high-degree vessels during neurovascular responses.
Vessel radius adjustments during optogenetic stimulation show 24 ± 28% variation relative to resting diameter, with dilations averaging 16.1 ± 14.3 µm and constrictions averaging 21.9 ± 14.6 µm relative to nearby neurons.
Wild-type C57BL/6J mice (n=4) show no statistically distinguishable capillary radius distributions following blue (458 nm) versus green (552 nm) photostimulation, confirming that vascular responses in Thy1-ChR2 mice are ChR2-specific and not attributable to photothermal or non-specific light effects.
Single-vessel measurements cannot predict network-level blood flow modulation: within-vessel heterogeneity, neuron-relative spatial gradients, assortativity rise, and efficiency changes together require a network-level treatment.
▸Dissociations
At 1.1 mW/mm² 458 nm stimulation, a sample artery dilated 1.33±0.86 µm (p<1e-4) and a sample capillary dilated 0.42±0.39 µm (p<1e-4), while a sample venule showed no significant radius change (p=0.22), demonstrating vessel-type heterogeneity in optogenetic neurovascular responses.
Vessel radius adjustments during optogenetic stimulation show 24 ± 28% variation relative to resting diameter, with dilations averaging 16.1 ± 14.3 µm and constrictions averaging 21.9 ± 14.6 µm relative to nearby neurons.
Constricting capillaries are located on average 37±179 µm deeper in cortex than dilating capillaries following 458 nm optogenetic stimulation at 4.3 mW/mm² (p<1e-4), with dilators tending toward the cortical surface across all stimulation conditions.
Following ChR2 optogenetic activation at 458 nm, capillary dilations occur on average 16.1±14.3 µm from the nearest labeled pyramidal neuron while constrictions occur 21.9±14.6 µm away (4.3 mW/mm²; p<1e-4), and this spatial segregation is absent under control 552 nm illumination.
UNETR ensemble segmentation shows significantly better HD95 surface distance than ilastik for both vessel and neuron channels (p<0.05 Wilcoxon signed-rank), while ilastik over-segments vessels with high recall (0.89±0.19) but low precision (0.37±0.33), evaluated on nine test images (507×507×250 µm) from six held-out mice.
The NOVAS3D deep learning segmentation pipeline (UNet/UNETR) achieves significantly higher Dice scores and precision/recall for volumetric vessel segmentation than the baseline ilastik classifier on the deposited two-photon microscopy test dataset.
▸Eliminations & validating controls
Capillary dilations following 458 nm ChR2 stimulation (0.90±0.93 µm at 1.1 mW/mm²; 0.90±0.78 µm at 4.3 mW/mm²) are significantly larger than those following 552 nm control illumination (0.58±0.92 µm; p<1e-4), confirming that the vascular responses are ChR2-mediated rather than photothermal artifacts.
Wild-type C57BL/6J mice (n=4) show no statistically distinguishable capillary radius distributions following blue (458 nm) versus green (552 nm) photostimulation, confirming that vascular responses in Thy1-ChR2 mice are ChR2-specific and not attributable to photothermal or non-specific light effects.
▸Synthesis claims
Single-vessel measurements cannot predict network-level blood flow modulation: within-vessel heterogeneity, neuron-relative spatial gradients, assortativity rise, and efficiency changes together require a network-level treatment.
▸Standalone empirical findings
The NOVAS3D segmentation model produces qualitatively reasonable vessel segmentations on out-of-distribution data including a different mouse strain (C57BL/6), a different species (Fischer rat), and a different microscope modality (light-sheet fluorescence microscopy, Miltenyi UltraMicroscope Blaze), without retraining.
▸Methodological warrants
The pipeline's boundary-detection radius estimation achieves R²=0.68 against ground-truth simulated radii across >100,000 simulations spanning 0.5×–2× dilation/constriction, and remains stable under Gaussian noise up to a standard deviation of 200 signal units (image range 0–1023 SU).
Rigid registration across time points and union of segmentation masks increases the number of identified vessel segments per FOV from 241±174 (single time point) to 412±281 (union across all time points; 507×507×250 µm, n=107 FOVs), while reducing mean squared error between acquisitions from 1306±747 to 0.008±0.003 signal units.
▸Scope qualifiers
All neurovascular coupling metrics are derived from a single integrated deep learning pipeline (segmentation → registration → graph analysis); the segmentation model requires GPU inference and was trained on data from one imaging preparation (Thy1-ChR2-YFP mice, 6-12 months), so generalization to other preparations is not demonstrated in this paper.
The NOVAS3D segmentation model was trained and quantitatively evaluated exclusively on data from Thy1-ChR2-YFP mice (line 18, 6–12 months, Texas Red vascular labeling, cranial window, isoflurane anesthesia); quantitative performance metrics are not reported for any other preparation, strain, species, or imaging modality.
Vessels are classified as responders if their radius change exceeds twice the baseline standard deviation (2×σ); this threshold is not sensitivity-tested in the main analysis, though an alternative 10% threshold is shown in Appendix 1—figure 14 with qualitatively similar results.
Pipeline-and-application scope: one DL stack trained on Thy1-ChR2-EYFP cranial-window 2PFM, benchmarked on six held-out mice, applied in vivo to n=17 Thy1-ChR2 + n=4 WT mice. OOD shown qualitatively only; no awake behaviour, no disease models, no alternative thresholds.
▸All claims (alphabetical)
- artery-dilates-venule-unchanged-at-low-power fig7A
- baseline-intra-vessel-radius-varies-24pct fig7
- blue-light-dilations-exceed-green-control fig8A
- capillary-efficiency-increases-4pct fig9C
- constrictions-deeper-than-dilations fig8E
- dilations-nearer-neurons-than-constrictions fig8D
- dl-model-scope-single-pipeline fig1 (architecture)
- hypothesis-dl-pipeline-enables-network-nvc hypothesis
- hypothesis-network-level-nvc-coordination hypothesis
- network-assortativity-increases-stimulation fig9B
- novas3d-generalizes-qualitatively-ood app1fig12, app1fig13
- novas3d-outperforms-ilastik fig3, fig4
- novas3d-single-preparation-scope fig1 (architecture); methods section
- prediction-pipeline-outperforms-baselines prediction
- prediction-pipeline-reveals-network-coordination prediction
- radius-estimation-r2-0p68 fig5
- registration-doubles-vessel-count
- responder-threshold-2sd-untested app1fig14
- scope-pipeline-and-application-paper scope
- synthesis-individual-vessel-measurements-insufficient synthesis (Discussion)
- unetr-outperforms-ilastik-hd95 fig3
- vessel-radius-heterogeneity-stimulation fig6
- wt-controls-no-blue-green-difference app1fig9
Abstract mapped to claims
The paper's abstract is shown with each sentence linked to the claim(s) it represents in the dependency graph. Hover or click a sentence to highlight the corresponding claim cards. Below: what the graph contains that the abstract leaves out, and vice versa.
1Functional hyperemia is a well-established hallmark of healthy brain function, whereby local brain blood flow adjusts in response to a change in the activity of the surrounding neurons. 2Although functional hyperemia has been extensively studied at the level of both tissue and individual vessels, vascular network-level coordination remains largely unknown. 3To bridge this gap, we developed a deep learning-based pipeline that uses two-photon fluorescence microscopy images of cerebral microcirculation to enable automated reconstruction and quantification of the geometric changes across the microvascular network, comprising hundreds of interconnected blood vessels, pre and post-activation of the neighboring neurons. 4The pipeline’s utility was demonstrated in the Thy1-ChR2 optogenetic mouse model, where we observed network-wide vessel radius changes to depend on the photostimulation intensity, with both dilations and constrictions occurring across the cortical depth, at an average of 16.1±14.3 μm (mean ± SD) away from the most proximal neuron for dilations; and at 21.9±14.6 μm away for constrictions. 5We observed a significant heterogeneity of the vascular radius changes within vessels, with radius adjustment varying by an average of 24 ± 28% of the resting diameter, likely reflecting the heterogeneity of the distribution of contractile cells on the vessel walls. 6A graph theory-based network analysis revealed that the assortativity of adjacent blood vessel responses rose by 152 ± 65% at 4.3 mW/mm² of blue photostimulation vs. the control, with a 4% median increase in the efficiency of the capillary networks during this level of blue photostimulation in relation to the baseline. 7Interrogating individual vessels is thus not sufficient to predict how the blood flow is modulated in the network. 8Our pipeline, enables tracking of the microvascular network geometry over time, relating caliber adjustments to vessel wall-associated cells’ state, and mapping network-level flow distribution impairments in experimental models of disease.
- H1.P1.4 unetr-outperforms-ilastik-hd95 fig3 UNETR ensemble segmentation shows significantly better HD95 surface distance than ilastik for both vessel and neuron channels (p<0.05 Wilcoxon signed-rank), while ilastik over-segments vessels with high recall (0.89±0.19) but low precision (0.37±0.33), evaluated on nine test images (507×507×250 µm) from six held-out mice.
- H1.P2.1 artery-dilates-venule-unchanged-at-low-power fig7A At 1.1 mW/mm² 458 nm stimulation, a sample artery dilated 1.33±0.86 µm (p<1e-4) and a sample capillary dilated 0.42±0.39 µm (p<1e-4), while a sample venule showed no significant radius change (p=0.22), demonstrating vessel-type heterogeneity in optogenetic neurovascular responses.
- H1.P2.9 wt-controls-no-blue-green-difference app1fig9 Wild-type C57BL/6J mice (n=4) show no statistically distinguishable capillary radius distributions following blue (458 nm) versus green (552 nm) photostimulation, confirming that vascular responses in Thy1-ChR2 mice are ChR2-specific and not attributable to photothermal or non-specific light effects.
- E1 novas3d-generalizes-qualitatively-ood app1fig12, app1fig13 The NOVAS3D segmentation model produces qualitatively reasonable vessel segmentations on out-of-distribution data including a different mouse strain (C57BL/6), a different species (Fischer rat), and a different microscope modality (light-sheet fluorescence microscopy, Miltenyi UltraMicroscope Blaze), without retraining.
- Sc3 responder-threshold-2sd-untested app1fig14 Vessels are classified as responders if their radius change exceeds twice the baseline standard deviation (2×σ); this threshold is not sensitivity-tested in the main analysis, though an alternative 10% threshold is shown in Appendix 1—figure 14 with qualitatively similar results.