Pith. sign in

REVIEW 4 major objections 6 minor 12 references

Automated anatomy-based post-processing reduces false positives and improved interpretability of deep learning intracranial aneurysm detection

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An automated anatomy-based filter that compares vein and artery overlap removes most false positives from two deep-learning aneurysm detectors while preserving all true positives.

desk verdict A useful, narrow post-hoc FP-filter paper with a genuinely informative FP taxonomy, but the headline zero-TP-loss claim is an in-sample result until independently validated. read the letter →

arxiv 2507.00832 v1 pith:SEASBCYE submitted 2025-07-01 eess.IV cs.AIcs.CV

Jisoo Kim , Chu-Hsuan Lin , Alberto Ceballos-Arroyo , Ping Liu , Huaizu Jiang , Shrikanth Yadav , Qi Wan , Lei Qin
show 1 more author
Geoffrey S Young
This is my paper · ORCID
classification eess.IVcs.AIcs.CV
keywords intracranialaneurysmdetectionfalse-positivereductionpost-processinganatomy-basedsegmentationCTangiographydeeplearningvein-arteryoverlapinterpretableAI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Deep-learning models that flag intracranial aneurysms on CT angiography are held back from clinical use by false positives, because every extra alert costs radiologist time and trust. This paper tests whether a fully automated post-processing step, built from anatomical segmentation masks rather than model retraining, can strip those false alarms out. On two detectors, the best filter removes any output outside the brain or whose overlap with a vein mask exceeds its overlap with an artery mask, eliminating 70.6% of false positives for one model and 51.6% for the other while the number of true positives stays exactly the same. The paper argues that false-positive reduction can be decoupled from the black-box model and grounded in interpretable anatomy, which also makes it easier to see why a given detection was removed.

What carries the argument

The load-bearing mechanism is the vein-more-than-artery overlap rule, combined with a brain mask. Each detector output is a 3D bounding box; the rule keeps a box only if its overlap with the arterial segmentation mask is at least as large as its overlap with the venous segmentation mask, and only if the box lies within a dilated brain mask with the cavernous venous sinus region added to include the skull base. The cavernous venous sinus is subtracted from the vein mask beforehand because it spatially overlaps common aneurysm locations. This one comparison encodes the anatomic fact that aneurysms arise from arteries: a true positive should look arterial, while the dominant false-positive classes—veins, extracranial vessels, and vein-adjacent tissue such as the choroid plexus—look venous or extracranial. It is the combination of the brain filter for extracranial false positives and the vein-versus-artery comparison for venous false positives that carries the argument.

What would settle it

Run Method 5 on a larger independent CTA cohort and count how many ground-truth aneurysm bounding boxes are removed; the paper's claim predicts zero, so a single deleted true positive, especially near the cavernous sinus or skull base, would refute it.

Watch

Extended reading notes

Core claim

Using 143 held-out CTAs with 218 annotated aneurysms, the paper shows that the false positives produced by CPM-Net and a deformable 3D CNN-transformer hybrid are not random noise: most sit in identifiable anatomic compartments, including extracranial tissue, veins such as the vein of Galen and extracranial veins, arterial branchpoints and infundibula, and non-vascular structures such as the choroid plexus. The post-processor applies four segmentation masks—brain, artery, vein, and cavernous venous sinus—to detector output boxes. Removing everything outside the dilated brain mask eliminates all extracranial false positives on both models. Removing boxes whose vein overlap exceeds artery overlap, after subtracting the cavernous sinus from the vein mask, removes the bulk of venous false positives while preserving every true positive: CPM-Net goes from 126 to 37 false positives and 3D-CNN-TR from 182 to 88, with true positives unchanged at 139 and 179. The paper interprets this as evidence that anatomy-grounded, interpretable post-processing can replace or complement threshold tuning as a false-positive reduction strategy.

Load-bearing premise

The no-true-positive-loss claim rests on the assumption that the automatic vein segmentation, after the cavernous venous sinus is subtracted, never overlaps a real aneurysm more than the artery segmentation does; if the vein mask bleeds into arterial lumen near the skull base, true positives would be deleted in a broader cohort.

Editorial extensions

If this is right

  • At the fixed confidence threshold of 0.8, the best filter cuts FP/case from 0.88 to 0.26 for CPM-Net and from 1.27 to 0.62 for 3D-CNN-TR, with true-positive counts unchanged at 139 and 179.
  • Brain-mask filtering alone removes all extracranial false positives, 27 of 27 for CPM-Net and 77 of 77 for 3D-CNN-TR, without removing a single true positive.
  • Venous false positives are the largest removable class, with 94.4% of CPM-Net's and 92.5% of 3D-CNN-TR's venous false positives removed by the vein mask.
  • Because the filter is applied after detection and uses only segmentation masks, it can in principle be attached to any existing detector without retraining.
  • The accompanying false-positive taxonomy, divided into extracranial, venous, arterial branchpoint, and non-vascular tissue categories, offers a map for designing further post-processing rules.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the same looks-more-venous-than-arterial overlap test could be transferred to other arterial targets such as arteriovenous malformations, dural fistulas, or venous sinus thrombosis, where the vessel of origin is known.
  • Beyond the paper, the unresolved false-positive class of normal arterial branchpoints and infundibula suggests that overlap-based masks have a natural limit, and that geometric vessel-shape criteria, such as saccular versus tubular contour analysis, would be the next logical filter to test.
  • Beyond the paper, because the method is applied at a fixed threshold, combining it with threshold tuning or an ensemble of detectors would likely push FP/case lower still, a combination the paper does not report.
  • Beyond the paper, a practical clinical check is whether the segmentation masks remain as reliable on routine CTA as on the dynamic CTA data used to develop the artery-vein segmentation algorithm; if not, the no-true-positive-loss result may shrink in broader deployment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes an automated anatomy-based post-processing method to reduce false positives in deep learning-based intracranial aneurysm detection on CTA. The method applies brain, artery, vein, and cavernous venous sinus (CVS) masks to the bounding-box outputs of two detectors (CPM-Net and a 3D-CNN-transformer hybrid, 3D-CNN-TR) and evaluates five removal rules on 143 held-out CTAs with 218 aneurysms. The authors report that the best rule (brain mask plus 'vein overlap greater than artery overlap', method 5) reduces false positives from 126 to 37 for CPM-Net and from 182 to 88 for 3D-CNN-TR, without reducing true positives, lowering FP/case from 0.88 to 0.26 and from 1.27 to 0.62 respectively.

Significance. If these results hold, the proposed post-processing is a simple, computationally cheap, and model-agnostic step that could meaningfully improve the clinical usability of DL-based aneurysm detectors. The systematic categorization of false positives by anatomic structure is a useful contribution to understanding detector failure modes. However, the central 'no true positive loss' claim is currently supported only by an in-sample estimate: the post-processing rules and mask parameters were selected after inspecting the FP/TP outputs on the same test set, and one of the two models was previously evaluated on this same cohort. The result is therefore promising but not yet established.

major comments (4)
  1. [Discussion (first paragraph)] The post-processing rules and parameters were chosen after inspecting FP/TP outputs on the same 143-CT test set used for the reported evaluation: the Discussion states 'Based on this FP analysis, we devised post-processing methods.' The 3.6 mm brain-mask dilation, the 3.2 mm CVS expansion, the vein-over-artery criterion, and the CVS subtraction are all effectively tuned on the test set. Consequently, the FP reductions in Table 2 are in-sample fitted estimates, and the claim that method 5 removes FPs 'without reducing TP' is not an out-of-sample result. For 3D-CNN-TR, this test set is also the same cohort used in its prior development paper (ref 2), further reducing independence. To establish the central claim, the authors should validate the pre-specified method on an external dataset, or at minimum use a split of the current data into development and validation subsets, and report the variability of the FP reduction.
  2. [Table 2] Table 2 is internally inconsistent with respect to TP and FN counts. For a fixed set of 218 ground-truth aneurysms, TP + FN must equal 218 in every row. The table lists CPM-Net method 4 as TP=129, FN=79 (sum 208) and 3D-CNN-TR method 4 as TP=169, FN=39 (sum 208). Since method 4 includes method 2, which the text states removes 10 TPs per model, the FN entries should be 89 and 49, respectively. Please correct the table and ensure the text and abstract do not carry the inconsistent numbers.
  3. [Results and Table 2] All results are reported at a single confidence threshold of 0.8 (Table 2 caption). The claim that method 5 reduces FPs without reducing TPs is therefore specific to that operating point; at other thresholds, the overlap between TP and vein masks may differ, so the method could remove TPs or fail to remove FPs. Please report sensitivity and FP/case across the full threshold range, or at least a few thresholds, and provide confidence intervals for the FP reduction proportions (e.g., Wilson intervals) to reflect sampling variability.
  4. [Methods (Artery and vein mask)] The artery-vein segmentation tool (ref 11) was developed on 4D dynamic CTA, and the paper does not quantify its accuracy on the standard CTA test images. The CVS subtraction is a manual atlas-based step with a 3.2 mm expansion, and the brain mask uses a 3.6 mm dilation; these parameters appear to have been selected on the same test set. If the vein mask bleeds into arterial lumen or the CVS subtraction is imperfect, true aneurysms in the skull base could be removed in other cohorts. Please report segmentation accuracy on a subset of the test data and perform a sensitivity analysis on the overlap ratio and dilation/expansion parameters.
minor comments (6)
  1. [Title] The title should read '...reduces false positives and improves interpretability...' for grammatical parallelism.
  2. [Table 2 caption] The phrase 'confidence interval of 0.8' should be 'confidence threshold of 0.8.'
  3. [Results] The statement '27/99 (27.3%) CPM-Net FP' is arithmetically incorrect for the extracranial fraction; Table 3 reports 27 extracranial of 126 total FPs, which is 21.4%. The 3D-CNN-TR figure (77/182, 42.3%) is correct.
  4. [Author affiliations] The affiliation line lists 'Washing University in St. Louis'—should be 'Washington University in St. Louis.'
  5. [Discussion] The typo 'dters radiologists' should be 'deters radiologists,' and 'cylindical' should be 'cylindrical.'
  6. [Abstract] The abbreviation CVS is used without spelling it out in the abstract; please define it at first use.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline FP-reduction and zero-TP-loss results are in-sample estimates because the post-processing rules and mask parameters were selected after analyzing the same 143-CTA evaluation set.

  1. fitted input called prediction [Discussion, paragraph 2; Methods, Post-processing (method 5 = methods 1 and 3); Results, Table 2]
    "To design a rational hybrid method for improving DL aneurysm detection on CTA, we performed a thorough analysis of all FP cases from two of the highest-performing intracranial aneurysm detection models. Based on this FP analysis, we devised post-processing methods that were effective in markedly reducing the FP rate at minimal computational cost and no additional user time cost."

    The five post-processing schemes, the CVS subtraction, and the mask expansions (3.2 mm for CVS, 3.6 mm for brain) were chosen after inspecting the FP and TP outcomes on the same 143-CTA evaluation set used for the reported results. Method 5 was constructed from Method 1 (brain mask) and Method 3 (vein-overlap greater than artery-overlap) precisely because these were observed to remove many FPs while preserving all TPs on that set. The reported reductions (CPM-Net 89/126; 3D-CNN-TR 94/182) and the central 'without reducing TP' claim are therefore in-sample summaries of the data used to select the rules, not independent predictions on new CTAs.

full rationale

The paper's numerical reporting is internally consistent and the anatomy-based idea is plausible, but the strongest claim that method 5 removes 70.6% and 51.6% of FPs without losing any TP is not an independent test. The authors explicitly say they devised the post-processing methods based on a thorough analysis of all FP cases on the evaluation set, and the mask dilation amounts, CVS subtraction, and vein-versus-artery overlap rule were all selected with knowledge of which outputs were FP and which were TP on those same 143 CTAs. Consequently, the zero-TP-loss result is an in-sample optimum rather than a prospective prediction, and it does not by itself establish that the method will generalize to new data or to different aneurysm locations. The self-citation to the authors' artery-vein segmentation tool (ref. 11) is a dependency, but it is not the main circular step because the masks are applied and visualized as tools rather than used as the sole proof of the central claim. No equation-level circularity is present, but the central validation claim reduces in part to the data used to tune the post-processing, so a moderate circularity score is warranted.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method depends on brain, artery, vein, and CVS masks being correct at aneurysm-bearing locations, on manual review labels, and on a single chosen threshold. Most parameters are hand-picked rather than fitted, but they are not justified by separate data. The post-processing rule itself was selected using the test set.

free parameters (4)
  • Detection confidence threshold = 0.8
    Both models' outputs were thresholded at 0.8; FP/TP counts and all post-processing results depend on this choice.
  • Brain mask dilation = 3.6 mm (9 voxels)
    Dilation ensures inclusion of intracranial structures but can alter which detections are considered extracranial; hand-chosen.
  • CVS region box expansion = 3.2 mm (8 voxels)
    Expansion of atlas-registered CVS box before intersecting with venous mask; hand-chosen and affects which detections are protected from vein-mask removal.
  • Vein-over-artery overlap criterion = vein voxels > artery voxels within bounding box
    Method 3/5 removal rule; no threshold calibration or independent validation provided.
assumptions (4)
  • domain assumption The nnUNet artery-vein segmentation from dynamic CTA (ref 11) correctly separates arteries and veins when applied to single-phase CTA.
    The entire vein-mask and vein-over-artery logic depends on this; no validation of the masks on the test set is reported.
  • domain assumption TotalSegmentator brain masks, dilated 3.6 mm and combined with the CVS box, reliably separate intracranial from extracranial structures.
    Method 1/4/5 remove all extracranial detections; if the mask mislabels skull base structures, true positives could be removed.
  • domain assumption The atlas-registered CVS mask correctly localizes the cavernous venous sinus after affine registration and expansion.
    CVS subtraction is meant to protect true aneurysms near the CVS; registration errors would undermine both TP preservation and FP removal.
  • domain assumption Manual neuroradiologist review of bounding boxes is an accurate reference for TP/FP labels.
    All quantitative conclusions are conditioned on the correctness of the manual annotations and the 'centered around the annotated aneurysm' criterion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automated anatomy-based post-processing reduces false positives and improved interpretability of deep learning intracranial aneurysm detection." pith.science (2026). https://pith.science/paper/SEASBCYE

@misc{pith2026250700832,
  author       = {Pith},
  title        = {Pith review of: Automated anatomy-based post-processing reduces false positives and improved interpretability of deep learning intracranial aneurysm detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SEASBCYE}},
  note         = {Machine review of arXiv:2507.00832}
}
read the original abstract

Introduction: Deep learning (DL) models can help detect intracranial aneurysms on CTA, but high false positive (FP) rates remain a barrier to clinical translation, despite improvement in model architectures and strategies like detection threshold tuning. We employed an automated, anatomy-based, heuristic-learning hybrid artery-vein segmentation post-processing method to further reduce FPs. Methods: Two DL models, CPM-Net and a deformable 3D convolutional neural network-transformer hybrid (3D-CNN-TR), were trained with 1,186 open-source CTAs (1,373 annotated aneurysms), and evaluated with 143 held-out private CTAs (218 annotated aneurysms). Brain, artery, vein, and cavernous venous sinus (CVS) segmentation masks were applied to remove possible FPs in the DL outputs that overlapped with: (1) brain mask; (2) vein mask; (3) vein more than artery masks; (4) brain plus vein mask; (5) brain plus vein more than artery masks. Results: CPM-Net yielded 139 true-positives (TP); 79 false-negative (FN); 126 FP. 3D-CNN-TR yielded 179 TP; 39 FN; 182 FP. FPs were commonly extracranial (CPM-Net 27.3%; 3D-CNN-TR 42.3%), venous (CPM-Net 56.3%; 3D-CNN-TR 29.1%), arterial (CPM-Net 11.9%; 3D-CNN-TR 53.3%), and non-vascular (CPM-Net 25.4%; 3D-CNN-TR 9.3%) structures. Method 5 performed best, reducing CPM-Net FP by 70.6% (89/126) and 3D-CNN-TR FP by 51.6% (94/182), without reducing TP, lowering the FP/case rate from 0.88 to 0.26 for CPM-NET, and from 1.27 to 0.62 for the 3D-CNN-TR. Conclusion: Anatomy-based, interpretable post-processing can improve DL-based aneurysm detection model performance. More broadly, automated, domain-informed, hybrid heuristic-learning processing holds promise for improving the performance and clinical acceptance of aneurysm detection models.

Figures

Figures reproduced from arXiv: 2507.00832 by the authors.

Figure 1
Figure 1. Arterial (yellow) and venous (green) segmentation masks of intra and extracranial [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Brain segmentation mask using TotalSegmentator, dilated 3.6mm in each direction [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

  1. [1]

    Short- and long-term mortality of subarachnoid hemorrhage according to hospital volume and severity using a nationwide multicenter registry study

    Park SW, Lee JY, Heo NH, Han JJ, Lee EC, Hong DY , et al. Short- and long-term mortality of subarachnoid hemorrhage according to hospital volume and severity using a nationwide multicenter registry study. Front Neurol. 2022;13:952794

    work page 2022

  2. [2]

    Vessel-aware aneurysm detection using multi-scale deformable 3D attention

    Ceballos-Arroyo AM, Nguyen HT, Zhu F, Yadav SM, Kim J, Qin L, et al. Vessel-aware aneurysm detection using multi-scale deformable 3D attention. Med Image Comput Comput Assist Interv. 2024 Oct;15005:754–65

    work page 2024

  3. [3]

    Detection of Intracranial Aneurysms Using Multiphase CT Angiography with a Deep Learning Model

    Wang J, Sun J, Xu J, Lu S, Wang H, Huang C, et al. Detection of Intracranial Aneurysms Using Multiphase CT Angiography with a Deep Learning Model. Acad Radiol. 2023 Nov;30(11):2477–86

    work page 2023

  4. [4]

    Detection of cerebral aneurysms using artificial intelligence: a systematic review and meta-analysis

    Din M, Agarwal S, Grzeda M, Wood DA, Modat M, Booth TC. Detection of cerebral aneurysms using artificial intelligence: a systematic review and meta-analysis. J Neurointerv Surg. 2023 Mar;15(3):262–71

    work page 2023

  5. [5]

    Effectiveness of tuning an artificial intelligence algorithm for cerebral aneurysm diagnosis: a study of 10,000 consecutive cases

    Kuwabara M, Ikawa F, Sakamoto S, Okazaki T, Ishii D, Hosogai M, et al. Effectiveness of tuning an artificial intelligence algorithm for cerebral aneurysm diagnosis: a study of 10,000 consecutive cases. Sci Rep. 2023 Sep 27;13(1):16202

    work page 2023

  6. [6]

    Deep Learning- Based Detection of Intracranial Aneurysms in 3D TOF-MRA

    Sichtermann T, Faron A, Sijben R, Teichert N, Freiherr J, Wiesmann M. Deep Learning- Based Detection of Intracranial Aneurysms in 3D TOF-MRA. AJNR Am J Neuroradiol. 2019 Jan;40(1):25–32

    work page 2019

  7. [7]

    Multidimensional Deep Learning Reduces False-Positives in the Automated Detection of Cerebral Aneurysms on Time-Of-Flight Magnetic Resonance Angiography: A Multi- Center Study

    Terasaki Y , Yokota H, Tashiro K, Maejima T, Takeuchi T, Kurosawa R, et al. Multidimensional Deep Learning Reduces False-Positives in the Automated Detection of Cerebral Aneurysms on Time-Of-Flight Magnetic Resonance Angiography: A Multi- Center Study. Front Neurol. 2021;12:742126

    work page 2021

  8. [8]

    CPM-Net: A 3D Center-Points Matching Network for Pulmonary Nodule Detection in CT Scans

    Song T, Chen J, Luo X, Huang Y, Liu X, Huang N, et al. CPM-Net: A 3D Center-Points Matching Network for Pulmonary Nodule Detection in CT Scans. In: Martel AL, Abolmaesumi P, Stoyanov D, Mateus D, Zuluaga MA, Zhou SK, et al., editors. Medical Image Computing and Computer Assisted Intervention – MICCAI 2020 [Internet]. Cham: Springer International Publishin...

    work page doi:10.1007/978-3-030-59725-2_53 2020

Show all 12 references
  1. [9]

    Toward human intervention-free clinical diagnosis of intracranial aneurysm via deep neural network

    Bo ZH, Qiao H, Tian C, Guo Y, Li W, Liang T, et al. Toward human intervention-free clinical diagnosis of intracranial aneurysm via deep neural network. Patterns (N Y). 2021 Feb 12;2(2):100197

    2021

  2. [10]

    3D Slicer as an image computing platform for the Quantitative Imaging Network

    Fedorov A, Beichel R, Kalpathy-Cramer J, Finet J, Fillion-Robin JC, Pujol S, et al. 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging. 2012 Nov;30(9):1323–41

    2012

  3. [11]

    Dynamic-Computed Tomography Angiography for Cerebral Vessel Templates and Segmentation [Internet]

    Yadav S, Kim J, Young G, Qin L. Dynamic-Computed Tomography Angiography for Cerebral Vessel Templates and Segmentation [Internet]. arXiv; 2025 [cited 2025 Jul 1]. Available from: http://arxiv.org/abs/2502.09893

    2025 arXiv

  4. [12]

    TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images

    Wasserthal J, Breit HC, Meyer MT, Pradella M, Hinck D, Sauter AW, et al. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiol Artif Intell. 2023 Sep;5(5):e230024

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.