Predictive modeling of hematoma expansion from non-contrast computed tomography in spontaneous intracerebral hemorrhage patients
- Department of Neurological Surgery, University of Virginia Health System, United States
- Department of Neurological Surgery, Thomas Jefferson University, United States
- Department of Radiology, University of Virginia Health System, United States
- Department of Electrical and Computer Engineering, North South University, Bangladesh
- Department of Radiology, Washington University in St. Louis, United States
- Department of Neurological Surgery, University of Louisville, United States
- Department of Biomedical Engineering, University of Virginia Health System, United States
- Department of Neurological Surgery, Tufts University Medical Center, United States
- Department of Neurosurgery, Catholic Health Initiative, United States
- Department of Biostatistics and Data Science, Wake Forest University School of Medicine, United States
- Department of Neurology, University of Buffalo, United States
- Department of Neurology, Westchester Medical Center, United States
- Department of Neurological Surgery, Vagelos College of Physicians and Surgeons, Columbia University, United States
- Department of Electrical and Computer Engineering, University of Virginia, United States
Figures
Figure 1
Example of the transport-based morphometry workflow.
(A) NCCT scan registration and segmentation. A population-based high-resolution NCCT template was used for NCCT registration prior to hematoma region segmentation. (B) Optimal transportation. Segmented hematoma regions, depicted as source images (I1, I2…), are transformed to the transport domain by pushing mass, represented as pixel intensity, from the source image to the reference image. This process computes optimal transportation maps, thereby representing the images as points on a high-dimensional Riemannian manifold. (C) Machine learning statistical analysis. In transport space, differences between given source images (I1, I2…) can be represented as a linear embedding of the difference between their computed transport maps. This permits effective application of statistical analysis methods to the high-dimensional data. (D) Discriminant image feature visualization. Representing data as points on a Riemannian manifold enables any point to be interrogated and inverted from the transport domain to the native image domain. This generates images of the discriminant features captured during statistical analysis. TBM is performed on volumetric NCCT images and two-dimensional slices are demonstrated for illustrative purposes. Abbreviations: PCA = principal components analysis, NCCT = non-contrast computed tomography, LDA = linear discriminant analysis.
Figure 2 with 2 supplements
Example of the preprocessing protocol.
(A) All NCCT scans were skull stripped and registered to a population-based high resolution NCCT template with dimensions of 256x256 x 256 and voxel spacings of 1x1 x 1 mm. (B) Example of a registered NCCT axial slice at presentation and (C) 24 hr in a patient with hematoma expansion. (F) The corresponding segmented and normalized hematoma image at presentation (G) and 24 hr for the same patient. (D) Example of a registered NCCT axial slice at presentation (E) and 24 hr in a patient without hematoma expansion. (H) The corresponding normalized hematoma image at presentation (I) and 24 hr in the same patient. Examples of segmented presentation non-contrast computed tomography (NCCT) hematoma images separated into groups of (J) hematoma expansion (≥6 mL hematoma volume increase at the 24 hr interval NCCT scan) and (K) no hematoma expansion (<6 mL hematoma volume increase at the 24 hr interval NCCT scan), demonstrating a lack of visually discernible difference between the two groups. Abbreviations: NCCT = non-contrast computed tomography.
Figure 2—figure supplement 1
Flow diagram illustrating the patient selection process.
Abbreviations: NCCT = Non-contrast computed tomography, DICOM = digital imaging and communications in medicine, ICH = intracerebral hemorrhage.
Figure 2—figure supplement 2
Two dimensional slice example illustrations of the template image.
The intrinsic mean template image used for the linear optimal transportation framework for the (A) original hematoma image data and (B) location-adjusted translated hematoma image data.
Figure 3 with 2 supplements
Results of the TBM model adjusted for location and clinical information in predicting 24 hr hematoma volume from the test dataset.
(A) Scatter plots showing the relationship between the hematoma image features in the test dataset projected onto the most correlated direction wcorr in transport space and change in hematoma volume from the presentation to the 24 hr NCCT scan. (B) Inverse transformations of three two-dimensional axial slice examples of the hematoma morphometric features found by the model to be associated with increasing growth, shown from left to right of the x-axis (C) Inverse transformations of the hematoma morphometric features overlaid onto the axial NCCT scan associated with least growth, left, and most growth, right. Abbreviations: NCCT = non-contrast computed tomography, TBM = transport-based morphometry, CC = correlation co-efficient σ=standard deviation of the pixel intensity distribution along wcorr.
Figure 3—figure supplement 1
Scatter plots showing the relationship between the hematoma image features in the internal validation cohort of the derivation dataset projected onto the most correlated direction in transport space and change in hematoma volume from the presentation to the 24 hr NCCT scan.
(A) TBM alone, (B) TBM adjusted for clinical information (C) TBM adjusted for location, (D) TBM adjusted for location and clinical information. Abbreviations: TBM = transport-based morphometry, CC = correlation coefficient, σ=standard deviation of the pixel intensity distribution along , δ=change, mL = milliliters.
Figure 3—figure supplement 2
Scatter plots showing the relationship between the hematoma image features in the training cohort of the derivation dataset projected onto the most correlated direction in transport space and change in hematoma volume from the presentation to the 24 hr NCCT scan.
(A) TBM alone, (B) TBM adjusted for clinical information, (C) TBM adjusted for location, (D) TBM adjusted for location and clinical information. Abbreviations: TBM = transport-based morphometry, CC = correlation coefficient, σ=standard deviation of the pixel intensity distribution along , δ=change, mL = milliliters.
Figure 4 with 7 supplements
Results of the TBM model adjusted for location and clinical information in predicting 24 hr hematoma expansion from the test dataset.
(A) Mean probability distributions of the hematoma image features in the test dataset projected onto the most discriminant direction w0 in transport space showing the degree of separation between the expansion (red) and no expansion (blue) groups by the learned pLDA classifier boundary. (B) AUROC analyses and corresponding 95% confidence intervals of the performance of the pLDA classifier in the test dataset for (C) Inverse transformations of three two-dimensional axial slice examples of the hematoma morphometric features found by the model to be associated with increasing likelihood of expansion, shown from left to right of the x-axis (D) Inverse transformations of the hematoma morphometric features overlaid onto the axial NCCT scan is least associated with expansion, left, and most associated with expansion, right. Abbreviations: NCCT = non-contrast computed tomography, TBM = transport-based morphometry, AUROC = area under the receiver operator curve, pLDA = penalized linear discriminant analysis, σ=standard deviation of the pixel intensity distribution along w0.
Figure 4—figure supplement 1
Mean probability distributions of the hematoma image features in the internal validation cohort of the derivation dataset projected onto the most discriminant direction in transport space showing the degree of separation between the expansion (red) and no expansion (blue) groups by the learned pLDA classifier boundary.
(A) TBM alone, (B) TBM adjusted for clinical information (C) TBM adjusted for location, (D) TBM adjusted for location and clinical information. Abbreviations: TBM = transport-based morphometry, pLDA = penalized linear discriminant analysis, σ=standard deviation of the pixel intensity distribution along .
Figure 4—figure supplement 2
AUROC analyses of the performance of the pLDA classifier.
Results from the internal validation cohort of the derivation dataset for (A) TBM alone, (B) TBM adjusted for clinical information, (C) TBM adjusted for location, (D) TBM adjusted for location and clinical information. Abbreviations: AUROC = Area Under the Receiver Operator Curve, ROC = Receiver Operator Curve, TBM = transport-based morphometry, pLDA = penalized linear discriminant analysis.
Figure 4—figure supplement 3
Mean probability distributions of the hematoma image features in the training cohort of the derivation dataset projected onto the most discriminant direction in transport space showing the degree of separation between the expansion (red) and no expansion (blue) groups by the learned pLDA classifier boundary.
(A) TBM alone, (B) TBM adjusted for clinical information, (C) TBM adjusted for location, (D) TBM adjusted for location and clinical information. Abbreviations: TBM = transport-based morphometry, pLDA = penalized linear discriminant analysis, σ=standard deviation of the pixel intensity distribution along .
Figure 4—figure supplement 4
AUROC analyses of the performance of the pLDA classifier.
In the training cohort of the derivation dataset for (A) TBM alone, (B) TBM adjusted for clinical information, (C) TBM adjusted for location, (D) TBM adjusted for location and clinical information. Abbreviations: AUROC = Area Under the Receiver Operator Curve, ROC = Receiver Operator Curve, TBM = transport-based morphometry, pLDA = penalized linear discriminant analysis.
Figure 4—figure supplement 5
Mean probability distributions of non-contrast computed tomography hematoma features in the internal validation cohort of the derivation dataset showing the degree of separation between the expansion (red) and no expansion (blue) groups.
(A) Initial hematoma volume, (B) Density heterogeneity, (C) Shape eccentricity, and (D) Image intensity distribution. Abbreviations: %=percent, mL = mililiters, HU = Hounsfield Units.
Figure 4—figure supplement 6
Mean probability distributions of non-contrast computed tomography hematoma features in the training cohort of the derivation dataset showing the degree of separation between the expansion (red) and no expansion (blue) groups.
(A) Initial hematoma volume, (B) Density heterogeneity, (C) Shape eccentricity, and (D) Image intensity distribution. Abbreviations: %=percent, mL = mililiters, HU = Hounsfield Units.
Figure 4—figure supplement 7
AUROC analyses of the performance of each of the visually identified image features as independent predictors of hematoma expansion with comparison to the final TBM model.
Features that were independent predictors of hematoma expansion in the (A) training cohort of the derivation dataset, (B) internal validation cohort of the training dataset and (C) external validation cohort of the test dataset. Abbreviations: AUROC = Area Under the Receiver Operator Curve, ROC = Receiver Operator Curve, TBM = transport-based morphometry.
Figure 5 with 1 supplement
Independent effects of hematoma location as a predictor of 24 hr hematoma expansion.
Two-dimensional examples of inverse transformations overlaid onto NCCT scans in the axial (top row), sagittal (second row), and coronal (third row) planes showing from left to right of the x-axis the hematoma morphometric features and location direction found by the TBM model to be associated with increasing likelihood of expansion. Abbreviations: NCCT = non-contrast computed tomography, TBM = transport-based morphometry, σ=standard deviation of the pixel intensity distribution along w0.
Figure 5—figure supplement 1
AUROC analyses of hematoma location as an independent predictor of hematoma expansion.
Results for (A) the internal validation cohort and (B) the training cohort of the derivation dataset. Mean probability distributions of hematoma location projected onto the most discriminant direction in transport space showing the degree of separation between the expansion (red) and no expansion (blue) groups by the learned pLDA classifier boundary in (C) the testing dataset and (D) the training dataset. Abbreviations: AUROC = Area Under the Receiver Operator Curve, ROC = Receiver Operator Curve, pLDA = penalized linear discriminant analysis, σ=standard deviation of the pixel intensity distribution along .
Figure 6 with 3 supplements
Comparisons of the performance of existing NCCT hematoma expansion prediction scores with comparison to the final TBM model adjusted for location and clinical information in the test dataset.
Hematoma expansion was defined as ≥6 mL hematoma volume increase from the presentation to the 24±6 hr NCCT scan. Abbreviations: AUROC = Area Under the Receiver Operator Curve, ROC = Receiver Operator Curve, TBM = transport-based morphometry, NCCT = non-contrast computed tomography, HEAVN = Heavn score, Brain = Brain score, HEP = Hematoma expansion prediction score, Pt = 10-point score, BAT = BAT score, TBM = transport-based morphometry.
Figure 6—figure supplement 1
Comparisons of the performance between existing NCCT hematoma expansion prediction scores and the final TBM model adjusted for location and clinical information in the (A) internal validation cohort and (B) training cohort of the derivation dataset.
Hematoma expansion was defined as ≥6 mL hematoma volume increase from the presentation to the 24±6 hr NCCT scan. Abbreviations: AUROC = Area Under the Receiver Operator Curve, ROC = Receiver Operator Curve, TBM = transport-based morphometry, NCCT = non-contrast computed tomography, Heavn = Heavn score, Brain = Brain score, NAG = NAG scale, PT = 10-point score, BAT = BAT score, TBM = transport-based morphometry.
Figure 6—figure supplement 2
Forest plots showing the mean correlation coefficient and corresponding 95% confidence intervals for each of the NCCT expansion prediction scores and 24 hr hematoma growth, measured as change in hematoma volume in milliliters from the presentation to the 24 hr NCCT scan, in the (A) internal validation cohort of the derivation dataset and (B) training cohort of the derivation dataset.
Abbreviations: NCCT = non-contrast computed tomography, Heavn = Heavn score, Brain = Brain score, NAG = NAG scale, PT = 10-point score, BAT = BAT score, TBM = transport-based morphometry.
Figure 6—figure supplement 3
Comparisons of the performance of alternate machine and deep learning methods and the final TBM model adjusted for location and clinical information in the external validation dataset.
Hematoma expansion was defined as ≥6 mL hematoma volume increase from the presentation to the 24±6 hr NCCT scan. Abbreviations: AUROC = Area Under the Receiver Operator Curve, ROC = Receiver Operator Curve, TBM = transport-based morphometry, NCCT = non-contrast computed tomography, KNN = K-nearest neighbors, SVM = support vector machine, Logistic = logistic regression, ResNet CNN = three-dimensional residual networks convolutional neural network.
Tables
Table 1
Comparison of demographic and clinical information between the expansion and no expansion groups in the derivation and test datasets.
Abbreviations: INR = international normalized ratio, S.D. = standard deviation, n=number, IVH = intraventricular hemorrhage, NCCT = non-contrast computed tomography, IQR = interquartile range. * on admission.
| Derivation dataset (n=170) | Test dataset (n=170) | |||||
|---|---|---|---|---|---|---|
| Expansion (n=56) | No expansion (n=114) | p-value | Expansion (n=56) | No expansion (n=114) | p-value | |
| Demographics | ||||||
| Age, years, mean ± S.D | 66.16±11.89 | 63.06±12.64 | 0.127 | 59.35±13.45 | 62.0±13.32 | 0.233 |
| Female, n (%) | 19 (33.9) | 45 (39.5) | 0.483 | 14 (25.9) | 35 (31.0) | 0.503 |
| Race/Ethnicity | 0.837 | |||||
| Black, n (%) | 11 (20.4) | 23 (20.3) | ||||
| Hispanic, n (%) | 22 (40.7) | 51 (45.1) | ||||
| White, n (%) | 21 (38.9) | 39 (34.5) | ||||
| Biochemistry | ||||||
| INR*, mean ± S.D. | 1.07±0.26 | 1.10±0.19 | 0.564 | 1.18±0.59 | 1.08±0.29 | 0.160 |
| Clinical parameters | ||||||
| SBP*, mean mmHg ± S.D. | 185±30 | 181±28 | 0.464 | 188±37 | 186±36 | 0.723 |
| Radiographic parameters | ||||||
| Time from symptom onset to NCCT, mean min ± S.D. | 104.13±37.7 | 113.58±41.3 | 0.151 | 505.11±527.27 | 934.35±1059.37 | 0.006 |
| Hematoma growth rate, mean mL/min ± S.D. | 0.426±0.279 | 0.281±0.375 | 0.012 | 0.152±0.202 | 0.004±0.018 | <0.001 |
| IVH score, median [IQR] | 0 [0–2] | 0 [0–1] | 0.405 | 0 [0–2] | 1 [0–2] | 0.107 |
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Predictive modeling of hematoma expansion from non-contrast computed tomography in spontaneous intracerebral hemorrhage patients
eLife 14:RP105782.
https://doi.org/10.7554/eLife.105782.3
