REVIEW 3 major objections 5 minor 11 references
Automated Registration of 3D Neurovascular Territory Atlas to 2D DSA for Targeted Quantitative Angiography Analysis
T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read Deformable registration of a 3D arterial territory atlas onto 2D DSA reaches median SSIM 0.81 across 2,247 DSAs, making territory-level quantitative angiography feasible during subarachnoid hemorrhage interventions.
desk verdict Feasibility report with a large dataset, but the headline SSIM is circular and 'deformable was essential' is asserted without a comparison. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is a 3D arterial-territory atlas in which each arterial region carries a numeric label, combined with a cone-beam projection that simulates the X-ray projection from the DSA's recorded geometry. The DSA side is summarized by a perfused-region mask built through temporal averaging, thresholding, small-component removal, morphological erosion, and hole filling. A two-stage registration — first affine, then B-spline — driven by a mutual information cost function aligns the projected atlas to the mask, with SSIM computed against the mask as the final overlay quality score.
What would settle it
Take 50 DSAs with independent expert manual tracing of arterial territories; if the B-spline overlay does not improve agreement with those tracings over the affine-only overlay, the central claim that deformable registration is essential would be falsified. A second check is to find a case where the perfused-region mask visibly misses a distal vessel branch: a high SSIM there would show the metric validating the mask rather than the anatomy.
Extended reading notes
Core claim
The central claim is that a two-stage registration — an affine transform initialized by the perfused-region mask, followed by a B-spline deformable transform — brings the projected 3D atlas into precise overlay with the DSA's perfused territories, and that this overlay can be segmented into named arterial territories automatically. The evidence is 2,247 DSAs with a median SSIM of 0.81 and a mean of 0.69±0.27, with a strongly left-skewed distribution; the authors interpret the skew as many near-perfect overlays plus a tail of poor ones. They also observe that exceptional overlays occur in higher proportions in anteroposterior views than lateral views. The paper is careful to state that quantitative territory analysis itself is not yet included; what is demonstrated is the registered, segmented overlay and saved transformation parameters that would support such analysis.
Load-bearing premise
The perfused-region mask derived from the DSA by averaging, thresholding, small-component removal, erosion, and hole filling is treated as ground truth for the perfused arterial territories, and SSIM is measured against that mask, so registration quality is only as good as the mask.
Editorial extensions
If this is right
- Once registered, each DSA can be segmented into named arterial territories, allowing quantitative angiography parameters to be read per territory rather than over the whole image.
- The saved affine and B-spline transformation parameters mean the atlas overlay can be recomputed for later analysis without re-running the registration.
- The full workflow from loading a DSA to a segmented overlay takes about 1 to 3 minutes, which is compatible with use during an endovascular intervention.
- The SSIM distribution is left-skewed with a median of 0.81, so most overlays are near the top of the quality range; the paper reports that exceptional cases appear more often in anteroposterior than in lateral views.
Reading between the lines
- Not stated in the paper: because the quality score is computed against the same perfused-region mask that guides the registration, the reported SSIM values may be optimistic for true anatomical overlap, and an independent manual or angiographic standard would be needed to know how much.
- A testable extension is to compare territory overlays from repeated injections of the same vessel in the same patient; consistency across repeats would give an artifact-free check of the registration that does not depend on the mask.
- The authors do not yet report territory-specific quantitative parameters; the saved transformation parameters make such territory-wise parametric maps a natural next step, and the same atlas-to-DSA pipeline could be adapted to other indications by changing the injection-site lookup table.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes an automated pipeline for registering a 3D neurovascular territory atlas (Brain Arterial Vascular Model) to 2D digital subtraction angiography (DSA) images, with the goal of overlaying arterial territory labels onto DSA for targeted quantitative angiography. The method comprises manual labeling of injection site and view, cone-beam projection of selected atlas regions using DICOM geometry, affine and B-spline registration of the projected atlas to a processed DSA mask, and evaluation via SSIM. On 2,247 DSA sequences, the authors report a mean SSIM of 0.69 ± 0.27 and a median of 0.81, and conclude that deformable registration was essential for precise overlay, enabling segmentation into arterial territories in 1–3 minutes. The paper claims that this approach enables targeted quantitative angiography analysis for SAH interventions.
Significance. If the registration accuracy were independently validated, the work would be a practical contribution to interventional neuroradiology: mapping a public 3D territory atlas onto routine 2D DSA in a few minutes could support territory-level visualization and future quantitative analysis. The scale of the evaluation (2,247 DSAs) is a notable strength, as is the transparent use of a public atlas and the explicit description of the imaging geometry. However, the validity of the central quantitative claim rests on the evaluation metric being computed against the same processed DSA mask used as the registration target, which is a circular measure; this, together with the unsupported assertion that deformable registration was 'essential,' limits the current evidence to feasibility. The paper is more a demonstration of an automated workflow than a validated clinical tool.
major comments (3)
- [Methods §2, Results §3] The evaluation is circular: the SSIM values reported in Results §3 are computed 'with reference to the processed DSA data' (Methods §2), which is the same temporally-averaged, thresholded, morphologically-cleaned mask that drives both the affine and B-spline registration. Measuring similarity to the registration target cannot distinguish a genuinely accurate anatomical overlay from one that has overfit the mask. The high median SSIM of 0.81 therefore partially reflects convergence to the target, not agreement with an independent standard. The authors should validate the overlays against an independent ground truth, for example manually delineated arterial territories by a clinician, or at minimum report expert visual scoring on a random sample of cases.
- [Results §3] The claim that 'affine transformations alone were insufficient' and that 'deformable registration was essential' is not supported by any quantitative comparison. The paper shows B-spline results but reports no affine-only SSIM values or statistical test comparing affine to deformable registration. Without these data, the necessity of the deformable step is an assertion rather than a finding. Please include affine-only SSIM (or a comparable metric) on the same 2,247 DSAs and a paired comparison.
- [Abstract, §4] There is a direct contradiction between the abstract's claim that the approach 'enabled the extraction of targeted quantitative angiography parameters' and §4's explicit statement that 'quantitative analysis was not included in this study.' The title also promises 'Targeted Quantitative Angiography Analysis,' but the paper only saves transformation parameters for future work. Either remove the quantitative-analysis claims from the abstract and title, or include the promised quantitative analysis (e.g., territory-specific wash-in/wash-out parameters) in the manuscript.
minor comments (5)
- [Abstract] The sentence 'This mask facilitated the initial coarse alignment of the projected 3D atlas to the DSA perfused territory deformable registration techniques' is grammatically incomplete; it appears to be missing a verb and continuity with the next sentence.
- [Methods §2] The first paragraph ends with 'excluded.at', which appears to be a typographical fragment; the sentence should be completed.
- [Methods §2] The mask generation parameters (threshold value, minimum component size, morphological structuring element, erosion kernel size) are not specified. Without these details the mask construction is not reproducible, and the sensitivity of SSIM to these choices is unknown.
- [Results §3] The histogram in Figure 2 shows a skewed distribution, but the paper does not report the number of cases with SSIM below, say, 0.5, nor does it analyze whether failures correlate with lateral vs. anteroposterior views, injection site, or image quality. Such a breakdown would strengthen the claims about 'higher proportions in anteroposterior views.'
- [General] The paper uses 'posterior' inconsistently in Methods §2: 'the projection view of the DSA data, either lateral or posterior' is likely meant to be 'lateral or anteroposterior'; using 'posterior' for a viewing direction is confusing because it is also an injection-site label.
Circularity Check
The headline SSIM is computed against the same processed-DSA mask that was the affine/B-spline registration target, so the reported overlay quality partly measures convergence to the fit target; no affine-only comparison supports the claim that deformable registration was essential.
-
fitted input called prediction
[Methods §2 (mask creation, affine and B-spline registration, SSIM evaluation); Results §3 (SSIM distribution).]
"Concurrently, a mask representing the perfused areas was created from the DSA sequence. This mask facilitated the initial coarse alignment of the projected 3D atlas to the DSA perfused territory through affine transformations. Further refinement of this alignment was achieved using deformable registration techniques, ensuring a precise overlay with the DSA’s perfused territories. ... Following the B-spline registration, performance of each overlay was quantified using the SSIM, with reference to the processed DSA data."
The 'processed DSA data' used as the SSIM reference is the same binary mask that served as the fixed registration target for the affine and B-spline optimizations, both driven by a mutual information cost function in SimpleITK. SSIM therefore measures post-optimization similarity to the optimization target itself, not agreement with an independent anatomical or expert standard. A high median SSIM (0.81) partly reflects convergence to the target mask and cannot distinguish a correct overlay from an optimizer that overfits the mask. The paper reports no affine-only SSIM baseline, so the conclusion that 'deformable registration was essential' is not supported by any independent quantitative comparison.
full rationale
The paper's central quantitative claim is the SSIM distribution (mean 0.69 ± 0.27, median 0.81) used to assert 'precise overlays' and that 'deformable registration was essential.' That metric is computed 'with reference to the processed DSA data,' which is the same temporally averaged, thresholded, morphologically cleaned mask that the affine and B-spline registrations were optimized to match. Consequently, the evaluation is self-referential: it reports goodness-of-fit to the registration target rather than validation against independent ground truth. The claim that deformable registration is essential is also not quantitatively supported, since no affine-only SSIM results are reported. These issues make the performance claim partially circular by construction. The atlas itself is external (Brain Arterial Vascular Model project, reference [11]), and no load-bearing self-citation chain appears, so the circularity is localized to the evaluation metric rather than the provenance of the atlas. The paper's own conclusion that 'quantitative analysis was not included in this study' further tempers the title's promise. Overall score 6 reflects one central evaluation that reduces to the fitting target.
Assumptions & free parameters
free parameters (2)
- DSA mask processing thresholds =
not reported
- Registration optimizer hyperparameters =
resolutions 20 (affine) and 16 (B-spline), max step 2
assumptions (3)
- domain assumption The Brain Arterial Vascular Model atlas (Liu et al. 2023) provides territory labels that correspond to actual DSA perfusion territories in SAH patients.
- domain assumption Temporal averaging and thresholding of DSA frames produce a mask that faithfully represents perfused arterial territories.
- domain assumption B-spline deformation of a generic atlas onto a patient's 2D projection preserves the clinical meaning of territory labels.
Cite this review
Pith. "Pith review of Automated Registration of 3D Neurovascular Territory Atlas to 2D DSA for Targeted Quantitative Angiography Analysis." pith.science (2026). https://pith.science/paper/FLPBHLOZ
@misc{pith2026241116637,
author = {Pith},
title = {Pith review of: Automated Registration of 3D Neurovascular Territory Atlas to 2D DSA for Targeted Quantitative Angiography Analysis},
year = {2026},
howpublished = {\url{https://pith.science/paper/FLPBHLOZ}},
note = {Machine review of arXiv:2411.16637}
}
read the original abstract
Subarachnoid hemorrhage (SAH), typically due to intracranial aneurysms, demands precise imaging for effective treatment. Digital Subtraction Angiography (DSA), despite being the gold standard, broadly visualizes cerebral blood flow, potentially masking key details in areas. This study introduces an approach integrating a 3D vascular atlas with 2D DSA images to allow targeted quantitative analysis in these crucial regions, thus enhancing diagnostic accuracy during interventions. Initially, DSA data was examined to ascertain the injection site. Following this, the appropriate viewing angle was determined to align accurately with the 3D vascular atlas. Utilizing this atlas, regions corresponding to the areas indicated as perfused were selected. Concurrently, a mask representing the perfused areas was created from the DSA sequence. This mask facilitated the initial coarse alignment of the projected 3D atlas to the DSA perfused territory deformable registration techniques, ensuring a precise overlay with the DSAs perfused territories. The performance of each overlay was measured using the Structural Similarity Index Measure (SSIM). The coregistration process revealed that deformable registrations was essential to achieve precise overlays of the 3D atlas projections with the 2D DSA perfused areas. This approach enabled the extraction of targeted quantitative angiography parameters, essential for detailed vascular assessment in subarachnoid hemorrhage cases. The integration of 3D atlas registration with 2D DSA projections facilitates a more precise and targeted diagnostic process for SAH during critical interventions. This image processing strategy enhances the visualization of affected arterial territories, potentially improving the accuracy of diagnostics and supporting better informed clinical decisions at the time of intervention
Figures
Reference graph
Works this paper leans on
-
[1]
G. Boulouis, C. Rodriguez -Regent, E. C. Rasolonjatovo, W. Ben Hassen, D. Trystram, M. Edjlali -Goujon, J. F. Meder, C. Oppenheim, and O. Naggara, “Unruptured intracranial aneurysms: An updated review of current concepts for risk factors, detection and management,” Rev Neurol (Paris), 173(9), 542-551 (2017)
work page 2017
-
[2]
The detection and management of unruptured intracranial aneurysms,
J. M. Wardlaw, and P. M. White, “The detection and management of unruptured intracranial aneurysms,” Brain, 123 ( Pt 2), 205-21 (2000)
work page 2000
-
[3]
R. Fogelholm, J. Hernesniemi, and M. Vapalahti, “Impact of early surgery on outcome after aneurysmal subarachnoid hemorrhage. A population-based study,” Stroke, 24(11), 1649-54 (1993)
work page 1993
-
[4]
Management of delayed cerebral ischemia after subarachnoid hemorrhage,
C. L. Francoeur, and S. A. Mayer, “Management of delayed cerebral ischemia after subarachnoid hemorrhage,” Crit Care, 20(1), 277 (2016)
work page 2016
-
[5]
Delayed Ischemic Neurologic Deficit after Aneurysmal Subarachnoid Hemorrhage,
V. N. Yamaki, D. D. Cavalcanti, and E. G. Figueiredo, “Delayed Ischemic Neurologic Deficit after Aneurysmal Subarachnoid Hemorrhage,” Asian J Neurosurg, 14(3), 641-647 (2019)
work page 2019
-
[6]
Cognitive and functional outcome after aneurysmal subarachnoid hemorrhage,
T. Al -Khindi, R. L. Macdonald, and T. A. Schweizer, “Cognitive and functional outcome after aneurysmal subarachnoid hemorrhage,” Stroke, 41(8), e519-36 (2010)
work page 2010
-
[7]
K. A. Williams, A. R. Podgorsak, M. M. S. Bhurwani, R. A. Rava, K. N. Sommer, and C. N. Ionita, “The Aneurysm Occlusion Assistant, an AI platform for real time surgical guidance of intracranial aneurysms,” Proc SPIE Int Soc Opt Eng, 11601, (2021)
work page 2021
-
[8]
M. M. Shiraz Bhurwani, M. Waqas, A. R. Podgorsak, K. A. Williams, J. M. Davies, K. Snyder, E. Levy, A. Siddiqui, and C. N. Ionita, “Feasibility study for use of angiographic parametric imaging and deep neural networks for intracranial aneurysm occlusion prediction,” J Neurointerv Surg, 12(7), 714-719 (2020)
work page 2020
Show all 11 references
-
[9]
Performance of angiographic parametric imaging in locating infarct core in large vessel occlusion acute ischemic stroke patients,
R. A. Rava, M. Mokin, K. V. Snyder, M. Waqas, A. H. Siddiqui, J. M. Davies, E. I. Levy, and C. N. Ionita, “Performance of angiographic parametric imaging in locating infarct core in large vessel occlusion acute ischemic stroke patients,” J Med Imaging (Bellingham), 7(1), 016001 (2020)
2020
-
[10]
Predicting the outcome of patients with subarachnoid hemorrhage using machine learning techniques,
P. de Toledo, P. M. Rios, A. Ledezma, A. Sanchis, J. F. Alen, and A. Lagares, “Predicting the outcome of patients with subarachnoid hemorrhage using machine learning techniques,” IEEE Trans Inf Technol Biomed, 13(5), 794- 801 (2009)
2009
-
[11]
Sci Data, 2023
Liu, C.F., et al., Digital 3D Brain MRI Arterial Territories Atlas. Sci Data, 2023. 10(1): p. 74
2023
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.