PaperlayerView exact revision on arXiv

Source-linked AI summary

Regional prediction of hematoma expansion from non-contrast computed tomography in spontaneous intracerebral hemorrhage patients

Natasha Ironside, Hongyu Kan, Kristofor Pas, Mohammad Shifat-E-Rabbi, Benjamin Newman, Shinjini Kundu, Kareem El Naamani, Tanvir Rizvi, Andrea Becerril-Gaitan, Harrison Snyder, Aarathi Manchikalapudi, Sarah Ratcliffe, Stacie L. Demel, Carl D. Langefeld, Daniel Woo, Stephan Mayer, E. Sander Connolly, Gustavo Rohde, for the ERICH Investigators, the VISTA-ICH collaboration

arXiv:2610.34645v1cs.CVq-bio.TO

TL;DR

Whether NCCT features predicting hematoma expansion vary by region matters because expansion and final volume relate to neurological outcomes. This study derived and validated regional TBM models from presentation scans; the combined validation model achieved a 0.787 AUROC. The results provide preliminary evidence for regional radiographic features, though prospective real-world validation is lacking.

  • Problem

    Whether NCCT features predicting hematoma expansion vary by hematoma region remains a research question, given the relationships between location, expansion, and neurological outcomes.

    Source: Introduction · p. 5

  • Method

    The study derived and externally validated region-stratified NCCT hematoma-expansion models using TBM representations and data from ERICH and VISTA.

    Source: Discussion · p. 20, Conclusions · p. 24

  • Results

    0.787 AUROC for the combined regional TBM model in validation, versus 0.775 for the feature-based model.

    Source: Discussion · p. 22

  • Takeaways & Limitations

    TBM visualized regional radiographic features of expansion, providing preliminary evidence that these features may inform prediction and understanding of expansion pathophysiology.

    Source: Conclusions · pp. 24–25

    The retrospective models lack prospective validation in a real-world setting, leaving their potential clinical applicability undetermined.

    Source: Discussion · p. 23

Abstract

from arXiv · show

Background The combination of hematoma location, expansion and final volume influence neurological outcome. The aim of this study was to derive and validate a quantitative and interpretable region-based NCCT image feature model for predicting hematoma expansion. Methods The Ethnic/Racial Variations of Intracerebral Hemorrhage(ERICH) study was used for derivation and the Virtual International Trials in Stroke Archive(VISTA) was used for validation. NCCT images at presentation and 24+/-6hours were pre-processed, segmented and categorized into lobar, basal ganglia or thalamic regions. Hematoma expansion was defined by region and according to its effect on clinical outcome. Transport-based morphometry (TBM) produced a unique transport-based representation from each NCCT hematoma image. Linear classifiers applied to transport representations were used to derive a predictive model. The regional TBM model's performance was compared to alternate methods using AUROC analyses. Results The derivation dataset included 574 patients, comprising 162 lobar, 257 basal ganglia and 155 thalamic hematomas. The validation dataset included 196 patients, comprising 31 lobar, 119 basal ganglia and 46 thalamic hematomas. A combined regional model predicted hematoma expansion in the validation dataset with an AUROC of 0.787. The principal image features which varied by location were size, heterogeneity, eccentricity and density distribution towards the hematoma periphery. The regional TBM model outperformed alternative clinician based and machine-learning methods. Conclusions TBM discriminated and visualized regional features of hematoma expansion from presentation NCCT scans. This provides preliminary evidence for regional radiographic features of hematoma expansion. It has the potential to improve hematoma expansion prediction and understanding of hematoma expansion pathophysiology.

Background

Hematoma expansion and final volume affect neurological outcomes, and location may shape their relationship. The study investigates whether region-specific NCCT features can predict expansion, using ERICH for derivation and VISTA for external validation.

Introduction

  • Location, initial and final hematoma volume, and scan timing may modify the relationship between expansion and neurological outcomes.Source: Introduction · p. 5
  • The study aims to identify, visualize, and evaluate region-based NCCT predictors while accounting for initial volume and time to the first scan.Source: Introduction · pp. 5–6

Study Population

  • ERICH supplied derivation cases, while VISTA supplied external validation cases; participants had supratentorial ICH and initial volumes of 2–60 mL.Source: Study Population · p. 6

Demographic, clinical and radiographic information

The study collected baseline clinical and radiographic measures and standardized NCCT images and hematoma segmentations for analysis.

Demographic, clinical and radiographic information

NCCT scan preprocessing, segmentation and volumetry

Outcomes

Expansion was assessed using region-specific clinical-outcome thresholds, with hematomas assigned to lobar, basal ganglia, or thalamic regions. TBM representations were reduced and combined with clinical modifiers to train regional classifiers.

Outcomes

  • Clinically significant expansion required volume growth of at least 6 mL or 33% and crossing the region-specific final-volume threshold associated with poor outcome.Poor outcome was defined by a modified Rankin Scale score of 3–6 at 90 days.Source: Outcomes · pp. 8–9

Defining hematoma location

  • Each hematoma was assigned to the lobar, basal ganglia, or thalamic region with the greatest overlap with a registered MRI atlas.Source: Defining hematoma location · p. 9

3D transport-based morphometry

Regional hematoma expansion model training

Regional hematoma expansion model testing

Regional classifiers were evaluated on validation data, and a patient-proportion-weighted combination produced an overall model. Inverse projections were used to visualize features distinguishing expansion from no expansion.

Regional hematoma expansion model testing

  • Regional validation performance was assessed by AUROC, with a derivation-set Youden threshold applied to validation data for classification accuracy measures.Sensitivity, specificity, and balanced accuracy were assessed.Source: Regional hematoma expansion model… · pp. 10–11
  • The overall TBM model averaged regional predicted probabilities weighted by each region's proportion of patients.Mean AUROC, sensitivity, specificity, and balanced accuracy assessed overall performance.Source: Regional hematoma expansion model… · p. 11

Visualization

  • Inverse projections visualized the direction and magnitude of feature differences between expansion and no-expansion groups in standard-deviation units.Source: Visualization · p. 11

Feature-based model

The study quantified and classified regional NCCT features, then compared the regional TBM model with clinician-based scores and alternative machine-learning methods. It assessed model performance with AUROC and classification metrics.

Feature-based model

  • Linear classifiers were trained on regional image-feature matrices, with validation performance assessed by AUROC and classification accuracy.The derivation-set Youden Index threshold was applied to validation data; accuracy used sensitivity, specificity, and balanced accuracy.Source: Feature-based model · pp. 11–12

Comparison to alternate clinician-based and machine learning methods

  • Clinician-based scores and six NCCT-based machine-learning models were evaluated against the regional TBM model using validation-set AUROC analyses.Two blinded trained raters computed BAT, BRAIN, HEAVN, NAG, and 10-point scores [8-10,32,35]; machine-learning candidates included SVM, KNN, logistic regression, random forest, 3D ResNet 18, and 3D DenseNet 121.Source: Comparison to alternate… · pp. 12–13

Statistical analyses

  • Baseline characteristics were compared using categorical or continuous-data tests as appropriate, with statistical significance set at p<0.05.Missing data were not imputed.Source: Statistical analyses · p. 13

Dataset composition

The study derived models in 574 ERICH patients and validated them in 196 VISTA patients. Hematoma expansion prevalence and regional composition differed across the cohorts and locations.

Dataset composition

  • Hematoma expansion was present in 17.1% of patients in the derivation dataset and 24.0% in the validation dataset.Source: Dataset composition · pp. 13–14
  • The validation cohort differed from derivation in mean age, systolic blood pressure, scan timing, hematoma growth rate, and intraventricular hemorrhage prevalence.Mean time to first NCCT was 971±1569 versus 109±39 minutes, and hematoma growth rate was 0.12±0.23 versus 0.22±0.20 mL/minute, respectively.Source: Dataset composition · p. 14

Region categories

  • Expansion rates varied by region: validation rates were 29.0% lobar, 27.8% basal ganglia, and 10.9% thalamic.Derivation rates were 22.2%, 19.1%, and 8.4%, respectively.Source: Region categories · pp. 14–15
  • Visual comparisons of segmented hematomas found no discernible image difference between expansion and no-expansion groups within regions.Source: Region categories · p. 15

Regional definitions of hematoma expansion

The study defined expansion using regional 24-hour hematoma-volume cutoffs linked to 90-day disability outcome, then tested regional TBM classifiers in validation data. Performance varied across the lobar, basal ganglia, and thalamic models.

Regional definitions of hematoma expansion

Regional TBM model discriminated 24-hour hematoma expansion from NCCT

  • Validation projections onto each region’s most discriminant transport direction showed statistically significant separation of mean expansion-probability distributions.Source: Regional TBM model discriminated… · p. 15

Lobar

  • The lobar TBM model achieved AUROC 0.773, balanced accuracy 72.7%, sensitivity 100.0%, and specificity 45.5% in validation.Initial hematoma volume and growth rate were included.Source: Lobar · pp. 15–16

Basal Ganglia

  • The basal ganglia TBM model achieved AUROC 0.739, balanced accuracy 68.2%, sensitivity 81.8%, and specificity 54.6% in validation.Initial hematoma volume and growth rate were included.Source: Basal Ganglia · p. 16

Thalamus

  • The thalamic TBM model achieved AUROC 0.820, balanced accuracy 63.4%, sensitivity 100.0%, and specificity 26.8% in validation.Initial hematoma volume and growth rate were included.Source: Thalamus · p. 16

Combined model

The combined regional TBM model achieved a mean validation AUROC of 0.787. Inverse transformations revealed location-specific hematoma features, which informed a regional feature-based modeling approach.

Combined model

  • The combined regional TBM model achieved mean validation AUROC 0.787, balanced accuracy 67.2%, sensitivity 88.2%, and specificity 47.6%.Initial hematoma volume and growth rate were included.Source: Combined model · p. 16

TBM discovered interpretable NCCT features of 24-hour hematoma expansion

  • Expansion-associated features included increasing size and peripheral density across regions, with heterogeneity in lobar and basal ganglia hematomas and eccentricity in basal ganglia and thalamic hematomas.The inverse-transformed features were plotted in standard-deviation units and visualized over NCCT images.Source: TBM discovered interpretable NCCT… · pp. 16–17

Regional feature-based model discriminated 24-hour hematoma expansion from NCCT

  • The feature-based model quantified TBM-identified features from registered, segmented, normalized NCCT images and combined them with hematoma growth rate.Source: Regional feature-based model… · p. 17

Lobar

  • The lobar feature-based model achieved validation AUROC 0.808, balanced accuracy 67.7%, sensitivity 64.1%, and specificity 72.7%.Source: Lobar · p. 17

Basal Ganglia

  • The basal ganglia feature-based model achieved validation AUROC 0.767, balanced accuracy 71.0%, sensitivity 90.9%, and specificity 51.2%.Source: Basal Ganglia · p. 18

Thalamus

Regional modeling identified location-dependent hematoma expansion features and produced predictive models that outperformed clinician-based and alternate machine-learning methods. The authors report that the findings are preliminary and that clinical applicability remains undetermined.

Thalamus

  • AUROC was 0.781 for the thalamus feature-based model in validation, with 70.5% balanced accuracy, 80.0% sensitivity, and 61.0% specificity.Source: Thalamus · p. 18

Discussion

  • Expansion-related features differed by location, and thalamic hematomas were less likely to expand than basal ganglia or lobar hematomas.The authors suggest these regional differences may reflect distinct biophysical mechanisms.Source: Discussion · p. 20
  • Regional classification achieved overall AUROC 0.79 versus 0.71 for the prior whole-brain model, possibly partly reflecting location-specific expansion definitions.Source: Discussion · p. 20
  • Inverting regional TBM representations made predictive image features interpretable, but proposed links between eccentricity and tissue pathways remain speculative.Basal ganglia expansion was predominantly oriented toward the lateral ventricle.Source: Discussion · p. 21
  • The regional feature-based model used fewer dimensions while achieving performance comparable to TBM, and the study remains hypothesis-generating rather than a readily implementable clinical tool.Source: Discussion · pp. 21–22
  • The combined TBM model achieved AUROC 0.787 versus 0.775 for the feature-based model; both outperformed clinician-based and alternate machine-learning methods.TBM had higher sensitivity, while the feature-based model had higher specificity.Source: based, and machine learning methods · p. 19, Discussion · p. 22

Conclusions

TBM discriminated and visualized regional features of 24-hour hematoma expansion from presentation NCCT scans, and the combined regional model achieved validation AUROC 0.79. The findings support preliminary evidence for regional radiographic features, with potential to improve prediction and understanding of expansion pathophysiology.

Conclusions

  • The combined regional TBM model achieved AUROC 0.79 in the validation dataset.Source: Conclusions · p. 24
  • The model's information was reproducible using feature-based statistical analysis, and it outperformed current clinician-based methods.Source: Conclusions · p. 25

Appendix 2

The figures show how patients were selected and hematomas were processed and assigned to regions, then illustrate regional TBM predictions and validation comparisons. They also depict learned features associated with expansion.

Figure 1

  • Figure 1 illustrates the patient selection process.The caption does not specify selection counts or criteria.Source: Figure 1 · p. 29

Figure 2

  • Figure 2 shows skull stripping, template registration, atlas-based regional outlines and assignment by greatest overlap; example scans lacked visually discernible group differences.Hematomas were allocated to lobar, basal ganglia or thalamic regions.Source: Figure 2 · p. 29

Figure 3

  • Figure 3 shows classifier separation between expansion and no-expansion groups and inverse-transformed regional features associated with increasing expansion likelihood.The displays cover lobar, basal ganglia and thalamic models.Source: Figure 3 · p. 29

Figure 4

  • Figure 4 presents validation-dataset AUROC analyses for lobar, basal ganglia, thalamic and overall regional models, adjusted for initial volume and growth rate.The caption does not provide AUROC values.Source: Figure 4 · p. 30

Figure 5

  • Figure 5 overlays inverse-transformed hematoma features on scans least and most associated with expansion for each regional TBM model.The examples cover lobar, basal ganglia and thalamic models.Source: Figure 5 · p. 30

Figure 6

  • Figure 6 compares existing NCCT expansion scores with the overall regional TBM and feature-based models in the validation dataset.The caption identifies the comparison but does not report the results or name a best-performing method.Source: Figure 6 · p. 30
Loading 2610.34645v1…