Semiautomated Volumetric Hemorrhage Quantification Using High-Frequency Ultrasound Following Spinal Cord Injury

Porcine model of SCI

This study included n = 17 Yucatan minipigs (Premier BioSource, Ramona, CA, USA), selected from the larger 7-day survival study (n = 30) (for an overview of the experimental design, see Fig. 1). A CONSORT-style flow diagram is provided as Supplementary Figure S1. All animals underwent a standardized contusion-compression SCI at the T10 spinal level as previously described. ³¹

OPEN IN VIEWER

FIG. 1.

Intraoperative high-frequency ultrasound (US) imaging of the spinal cord (SC) in a porcine model of SCI. Overview of our experimental design. All animals received a T10 contusion-compression SCI (red star). (A) Three-dimensional (3D) in vivo brightness-mode (B-mode) US scans were acquired at several time points (green circles): 0.5 h before injury (pre-SCI), 0.5 h post-injury (HPI), hourly up to 7 h, and 168 h post-injury (7 days). Image created using BioRender (https://www.biorender.com). (B) The US imaging system (Vevo 3100, FUJIFILM VisualSonics, Toronto, Canada) consists of the following components: (1) high-frequency transducer (30 MHz, MX400), (2) 3D motor (60 mm), (3) transducer mount, (4) touchscreen interface, and a (5) high-definition monitor. (C) Representative midsagittal images obtained pre-SCI and 0.5 h post-injury from one animal. (D) The gray-value range for intraparenchymal hemorrhage (IPH) on US was determined by analyzing the histogram within a region of interest (yellow box) where the hyperechoic signal on US overlapped with red blood cells (RBCs) visible on hematoxylin and eosin (H&E) staining. (E) Using the gray-value range determined in (D), the SC (in blue) and IPH (in red) were segmented using our semiautomated approach to create 3D masks. This allowed us to assess the extent of IPH pre-SCI and at 0.5, 2, 4, and 7 h post-injury. (F) To evaluate the performance of the semiautomated method, we compared IPH measurements derived from both manual and semiautomated segmentation, assessing inter-rater and intra-rater reliability. Imaging parameters: gain: 22 dB; dynamic range: 25 dB; axial slice thickness: 0.203 mm; scanning range: 40 mm. Image created using BioRender. AB, A. (2026). https://BioRender.com/rfcw0y0. SCI, spinal cord injury.

Anesthetic induction was accomplished via intramuscular Telazol^® (5 mg/kg) and xylazine (1 mg/kg). Following induction, animals were intubated and mechanically ventilated at 10–12 breaths/min with a tidal volume of 12–15 mL/kg. Isoflurane was used for induction (2–3%) and then immediately discontinued. It was subsequently reintroduced at 1% during cauterization, which lasted approximately 30 min to 1 h.

Throughout the experiment, animals received continuous intravenous infusions of propofol (5–9 mg/kg/h), fentanyl (4–12 µg/kg/h), ketamine (4–12 mg/kg/h), and midazolam (0.2 mg/kg/h, discontinued after SCI). Intravenous fluids were administered at 3–7 mL/kg/h using a combination of Isolyte^® and 5% dextrose in water with 0.9% NaCl; a subset of animals received 1.25% dextrose. Six min prior to SCI, a fentanyl bolus (7 µg/kg over 1 min) was administered, and isoflurane was reintroduced at 2% for the final 5 min before SCI and discontinued immediately after the contusion injury.

A dorsal midline skin incision was made in the thoracic region, and a laminectomy from T6 to T14 was performed and widened to expose the SC from T8 to T12. A contusive SCI (weight: 50 g; height: 16.73 cm) was delivered at T10 using a weight-drop impactor device. Following contusion, an additional 100 g weight was added to the impactor (150 g total) to sustain a 5-min compression period on the SC. Peak force, impact velocity, and maximum displacement were recorded and analyzed with LabVIEW software (National Instruments Corporation, Austin, TX, USA).

After the last post-injury intraoperative US scan (described in more detail below), the incision was closed, and animals recovered for 7 days. ³² Post-surgical pain management included fentanyl administered via continuous rate infusion (0.1–1.25 µg/kg/h) or a 12-µg patch delivering 0.41–0.55 µg/kg/h. Meloxicam (0.2 mg/kg) was given orally or subcutaneously daily for 3–4 days. Animals received either cefazolin (22 mg/kg) intravenously for 1 day or enrofloxacin (5 mg/kg) intravenously for 3 days.

At the 6-h (n = 3, Cohort 1) or 7-day post-SCI end-point (n = 14, Cohort 2), animals were euthanized using an overdose of pentobarbital (Euthanyl^®) administered intravenously (120 mg/kg) by a licensed animal care technician/veterinarian and confirmed by auscultation of the heart.

US morphometric measures of SC and CSF space

High-frequency US imaging was used to perform a morphometric assessment of the SC and surrounding cerebrospinal fluid (CSF)-containing dural sac (referred to as the subdural, intrathecal, or subarachnoid space). Axial US images acquired 30 min before SCI were manually analyzed using Vevo LAB software (www.visualsonics.com, version 5.5.1.6004-x64), as previously described, ²² to determine the following variables: SC_area = SC area in mm², DS_area = dural sac area in mm², SC_area:DS_area = SC area:dural sac area ratio, CSF_D = dorsal thickness of the CSF layer in mm, CSF_V = ventral thickness of the CSF layer in mm, and CSF_DV in mm = dorsal–ventral thickness of the CSF layer in mm.

Acquisition of intraoperative high-frequency US images

Intraoperative B-mode US imaging of the porcine SC was performed using a high-frequency 30-MHz transducer (depth: 20 mm; width: 13.36 mm; Vevo 3100, FUJIFILM VisualSonics, Toronto, Canada). On the day of the SCI procedure, the first US session took place 30 min before SCI, followed by hourly imaging for up to 7 h post-SCI. Seven days after SCI, a second surgery was performed to expose the SC for the final imaging session to observe IPH in the subacute phase after SCI. Anesthesia methods were consistent with those used on the SCI surgery day (see Supplementary Table S1 for details).

Registered veterinary technicians continuously monitored and adjusted anesthetic depth throughout the procedure using multiple physiological and clinical indicators, including jaw tone, heart rate (60–140 beats/min), and mean arterial pressure (>60 mmHg), to maintain physiological stability and an appropriate surgical plane of anesthesia in accordance with institutional animal care guidelines.

Cardiovascular and respiratory parameters were also recorded continuously during both surgical procedures (SCI surgery on Day 0 and sac surgery on Day 7) as part of the larger 7-day survival study. Although these parameters were not formally compared statistically between surgical procedures, physiological variables remained within institutional target ranges throughout both anesthetic sessions, and no major cardiovascular or respiratory instability requiring deviation from the anesthetic protocol was observed. Respiratory rate was controlled via mechanical ventilation (10–12 breaths/min) and was therefore not used as an indicator of anesthetic depth.

Sterile saline was used as the coupling agent for intraoperative US scanning. It was suctioned and replaced with fresh sterile saline before each scan to maintain acoustic coupling and minimize accumulation of blood or debris between imaging time points. To minimize respiratory artifacts, the animal’s respiration was temporarily paused at the peak of inspiration during image acquisition.

For each subject, three-dimensional (3D) B-mode scans were acquired before injury (30 min before SCI) and at multiple time points post-injury as part of the larger 7-day survival study. The present analysis focused on 0.5, 2, 4, and 7 h post-injury (HPI), selected to capture hyperacute IPH evolution. At each time point, scans were acquired using two different acquisition settings to evaluate the effect of gain and dynamic range on IPH quantification. Standard acquisition settings consisted of a gain of 22 dB and a dynamic range of 25 dB, whereas adjusted acquisition settings consisted of a gain of 18 dB and a dynamic range of 50 dB. Both datasets were analyzed separately. However, only the 4-h post-injury scans were used for direct comparison of IPH measurements between acquisition settings, as this time point served as the pre-treatment baseline IPH assessment in the broader study.

Initial scans were acquired using a gain of 22 dB and a dynamic range of 25 dB, which served as the standard in vivo acquisition settings for IPH assessment in this study and were consistent with the settings recommended by FUJIFILM VisualSonics at the time of system installation. Given the higher-contrast produced by these standard settings, the dorsolateral aspects of the white matter appeared brighter relative to the adjacent white matter. Therefore, additional images were acquired using adjusted settings (gain: 18 dB; dynamic range: 50 dB) to reduce overall image contrast and produce a more uniform echogenic appearance of the white matter. For both acquisition settings, the gray-value range corresponding to IPH was defined based on images acquired within the first 6 h post-SCI (n = 3, Cohort 1) and applied for semiautomated segmentation.

Using a 6.0 cm 3D motor (VisualSonics), serial 2D axial images were acquired with the scan centered at the T10 spinal level (i.e., site of impact). Localization was performed using US identification of the hypoechoic intervertebral disc as an anatomical landmark, enabling consistent rostral–caudal positioning across time points and animals. Images acquired pre-SCI, at 0.5 h post-SCI, and hourly thereafter up to 7 h were collected over a 40-mm scanning range with a step size of 0.203 mm, yielding 197 axial images per dataset and enabling capture of the full extent of IPH. The 0.203-mm increment was selected based on pilot testing to balance spatial resolution and 3D scan acquisition efficiency.

At 7 days post-SCI, the rostral–caudal extent of IPH extended beyond the initial 40-mm scanning range; a 63.4-mm scanning range was therefore implemented for this time point, generating 312 images per dataset. All resulting 3D reconstructed volumes were then exported from the Vevo 3100 US system for analysis.

Motion correction and image denoising

Motion correction was conducted using custom-built image stabilization software within the Vevo LAB US analysis platform (https://www.visualsonics.com), aimed at mitigating cord pulsation artifacts observed in the sagittal plane during 3D image reconstruction (Fig. 2A). Following motion correction, each 3D scan was converted from a cine loop to Digital Imaging and Communications in Medicine (DICOM) format using Vevo LAB. An in-house MATLAB script, adapted from the methodology presented by Coupé et al. (2009), ³³ was subsequently employed to perform image denoising (Fig. 2B). The resulting denoised DICOM images were then rescaled in 3D Slicer (www.slicer.org, version 5.2.2), where the image spacing values for all animals were adjusted to 0.025 mm in height, 0.025 mm in width, and 0.203 mm in slice thickness. These adjustments were made based on metadata extracted from the original DICOM files using MicroDicom DICOM Viewer (www.microdicom.com).

OPEN IN VIEWER

FIG. 2.

Quantifying IPH within the SC on high-frequency US images using a semiautomated approach. A semiautomated pipeline was developed to quantify IPH progression following SCI. (A) Motion artifact removal involved image stabilization of each 3D intraoperative US scan to mitigate cord pulsation artifacts in the sagittal plane (yellow triangles) using Vevo LAB software. Orientation: D = dorsal; V = ventral; R = rostral; C = caudal. (B) Next, image denoising was performed using an in-house MATLAB script. (C) In 3D Slicer, the 2D SC area (blue mask) was manually segmented in the axial plane at 5-mm intervals to isolate the SC from the pia mater. Inter-slice interpolation was automatically applied to create the 3D SC mask, which was then eroded with a margin size of 0.21 mm. (D) For IPH segmentation, global thresholding (gray-value range: 170–248) was applied within the eroded 3D SC mask, followed by manual corrections based on established rules (see Supplementary Fig. S2). (E) Using the reconstructed 3D models, IPH volume and surface area (SA) for both the SC and IPH were extracted. Imaging parameters: gain: 22 dB; dynamic range: 25 dB; axial slice thickness: 0.203 mm; scanning range: 40 mm. IPH, intraparenchymal hemorrhage; SC, spinal cord; SCI, spinal cord injury; US, ultrasound.

Segmentation of the SC and IPH and quantification

Utilizing denoised DICOM images, the SC parenchyma was manually segmented in the axial plane at 5-mm intervals across a 40-mm segment centered at the site of impact (Fig. 2C). In 3D Slicer, contour interpolation was applied to automatically bridge gaps between the nine segmented slices, creating a preliminary 3D SC mask. Pixels outside this mask (pia mater and surrounding tissue) were removed automatically by eroding the SC mask by a margin of 0.21 mm.

Global thresholding was used to create the initial 3D IPH mask (Fig. 2D) using a gray-value range of 170–248 for the standard US acquisition settings (gain: 22 dB, dynamic range: 25 dB) and 110–192 for the modified settings (gain: 18 dB, dynamic range: 50 dB).

To establish fixed setting-specific threshold ranges for semiautomated IPH segmentation, spatial and temporal histogram analyses were performed in ImageJ (https://imagej.net/ij/) using 8-bit resolution grayscale images (0–255, where 0 = black pixels and 255 = white). Cohort 1 animals (6-h post-SCI end-point) were used for IPH threshold range determination.

Nine US images from three animals (three images per animal), acquired at 6 h post-SCI using standard imaging settings (gain: 22 dB, dynamic range: 25 dB) and adjusted settings (gain: 18 dB, dynamic range: 50 dB), were analyzed. Regions of interest (ROIs) were defined within hyperechoic areas on US and verified visually against corresponding histology; however, spatial co-registration was not performed due to differences in resolution (Fig. 5A).

To evaluate the temporal consistency of the hyperechoic signal, ROIs were additionally tracked across serial images acquired at 0.5 h post-SCI and hourly from 1 to 6 h post-SCI. Histogram profiles were generated for each ROI to assess variability in gray-value distributions over time. The final threshold ranges (170–248 for Setting 1 and 110–192 for Setting 2) were defined separately for each acquisition setting, based on the minimum and maximum grayscale values observed across all corresponding histograms.

Although histological validation was performed only up to 6 h post-SCI, 7-hour US datasets were included to extend characterization of hyperacute IPH evolution across the full intraoperative imaging period. Because temporal histogram analyses demonstrated relative stability in gray-value distributions between 4 and 6 h post-SCI, the same threshold range was subsequently applied to the 7-h datasets.

Final IPH masks were created through manual refinement using rule-based criteria (see Supplementary Fig. S2). The first criterion excluded segmented pixels extending beyond the rostral–caudal boundaries of the lesion, identified in axial images showing visible gray matter structure (rostral–caudal trimming). The second criterion excluded pixels near the SC perimeter across all axial sections, particularly those related to the pia mater or with gray values similar to healthy white matter (axial trimming).

IPH outcome measurements

Using the final SC and IPH masks, we automatically calculated the SC and IPH area (mm²), volume (mm³), and surface area (SA; mm²) using 3D Slicer software (Fig. 2E). The change in IPH (delta IPH) was quantified by assessing the differences in volume, SA, or length (mm) measures between the initial and follow-up scans. The length of IPH in the rostral–caudal direction was assessed in the axial plane. Additionally, IPH was normalized relative to the SC volume over a 40-mm segment, computed by dividing the volume of the IPH by the volume of the SC and multiplying the result by 100. The rate of IPH expansion was calculated as the change in volume, percentage, SA, or length between the initial and follow-up US scans, divided by the interval elapsed between the two imaging time points.

Assessment of the semiautomated IPH quantification approach

To assess inter- and intra-rater reliability efficiently, a subset of five slices per animal (15 images total across three animals) was selected prior to manual segmentation. Slices were chosen by reviewing the axial image stack within ±5 mm of the impact site (where the contrast between IPH and non-hemorrhagic tissue is greatest) and selecting those that captured distinct morphological changes in IPH appearance along the rostral–caudal axis, rather than using fixed-interval sampling. This approach was used to avoid redundancy from adjacent slices with minimal visual differences in IPH distribution and to reduce potential bias from oversampling similar sections. Semiautomated segmentation was performed on the full dataset (197 images per animal). One month later, the same images were re-segmented for intra-rater reliability assessment.

Motion correction, image denoising, and rescaling were completed prior to analysis. Raters received fully preprocessed image volumes and were responsible only for semiautomated SC masking and IPH segmentation, ensuring that variability in segmentation was assessed independently of preprocessing. Standard US acquisition settings (gain: 22 dB, dynamic range: 25 dB) and the fixed, standardized global thresholding range (170–248) were applied consistently across all datasets.

To capture a range of US analysis experience during initial validation, the rater cohort included the rater who developed and optimized the semiautomated method (Rater 1). All raters underwent two days of training (2 h/day) and subsequently analyzed a separate training dataset over 5 days (∼1–2 h/day), using standardized procedures to minimize bias.

SC processing and histology

At the 6-h (n = 3, Cohort 1) or 7-day post-SCI (n = 14, Cohort 2) experimental end-points, a 60-mm SC segment surrounding the impact site was collected from each animal and fixed in 10% formalin at room temperature for 4 days. After 2 weeks of cryoprotection in sucrose solutions (12–30%), each segment was cut into 12-mm blocks using a 3D-printed mold and embedded in optimal cutting temperature compound (Tissue-Tek). Each 12-mm block was serially sectioned into 20-µm-thick axial sections using a cryostat (HM525 NX; Fisher Scientific, Ottawa, Ontario, Canada). Sections were distributed across 20 slides per set, with each slide containing sections spaced 400 µm apart. Sections were organized into five sequential sets per block. For each set, the middle three slides were selected for histological staining: one for hematoxylin and eosin (H&E), one for Prussian blue, and one for combined H&E + Prussian blue, resulting in a total of 15 stained slides per block. All remaining slides were stored at −80°C.

For H&E staining, slides were cleared in xylene, rehydrated through graded alcohols (100%, 95%, 70%), and water. They were stained with Gill’s Hematoxylin Solution No. 2 (3 min), differentiated in 0.5% hydrochloric acid (HCl, 10 sec), then stained with bluing reagent (15 sec). Eosin Y was applied (5 min) for cytoplasmic staining.

For Prussian blue staining, slides were cleared in xylene, rehydrated through graded alcohols (100%, 95%, 70%), and stained with an iron stain solution (2% potassium ferrocyanide solution combined with 2% HCl) for 30 min. After rinsing, a nuclear fast red counterstain was applied (5 min).

For combined H&E and Prussian blue staining, ³⁴ slides were cleared, rehydrated, and stained with the iron solution for 30 min. They were then stained with Gill’s Hematoxylin Solution No. 2 and differentiated in 0.5% HCl (10 sec each), and stained with bluing reagent (15 sec). Eosin Y was applied for 5 min. After dehydration and clearing, slides were mounted, coverslipped, and scanned at 20× (Aperio AT2 digital scanner).

Alignment between axial intraoperative US and histology images was achieved by visually matching hyperechoic regions observed on US with red blood cells (RBCs) on H&E staining and iron on Prussian blue staining, thereby ensuring consistency in their presence, shape, and location within the SC. To enhance the accuracy of image registration between in vivo and ex vivo images, minimal linear transformations (i.e., image scaling and rotation) were utilized to alter the SC shape of the histology images.

Data and statistical analysis

Data are presented as mean ± standard deviation (SD). All analyses were performed using GraphPad Prism 10, SPSS (version 29), and 3D Slicer (version 5.2.2), with significance set at p <0.05.

Friedman tests were used to evaluate significant differences in SC volume, IPH volume, 3D IPH SA, and IPH length measurements across multiple post-SCI time points. Changes in SC swelling (rate of SC volume change) and IPH expansion rate (rate of IPH volume, percentage, SA, and length change) were also analyzed. Dunn’s post hoc test adjusted for six planned comparisons. For the SC volume analyses, 10 planned comparisons were adjusted.

2D IPH area (mm²) values at 0.5, 2, 4, and 7 h post-SCI were log transformed after adding a small constant (0.01) to accommodate zero values. A mixed-effects model fit using restricted maximum likelihood (REML) was used to assess the main effects of time and location.

Two-tailed Wilcoxon signed-rank tests compared IPH volume and 3D IPH SA between images acquired using two different US acquisition settings.

Inter-rater agreement was also evaluated in 3D Slicer by mapping regions identified by all three raters, any two raters, or only one rater. From these segmentation agreement maps, the percentage of total 2D IPH area (mm²) or IPH volume (mm³) identified by all three raters was calculated. Dice similarity coefficient (DSC) scores were calculated for IPH area and volume overlap among raters.

Inter-rater reliability for IPH segmentation using manual and semiautomated methods was evaluated in SPSS using intraclass correlation coefficients (ICC) derived from a two-way mixed-effects model for absolute agreement, based on pooled area measurements from five images per animal across three animals. ICC values were also calculated using linear mixed-effects models to account for repeated measures. For inter-rater reliability, the model included animal, images nested within animals, and rater as random effects. For intra-rater reliability, rater was excluded from the model. ICC values >0.90 were considered indicative of excellent reliability, values between 0.75 and 0.90 were considered indicative of good reliability, and values between 0.50 and 0.75 were considered indicative of moderate reliability. ³⁵

Exploratory Spearman’s rank correlations (two-tailed) of pooled measurements assessed relationships among pre-injury spinal and CSF morphometrics, biomechanical injury parameters, and IPH, as well as between SC swelling and IPH, and between 2D IPH area measurements from manual and semiautomated methods for each rater (Raters 1–3). Exploratory Bland–Altman analyses of pooled measurements assessed agreement between segmentation methods, calculating the mean bias and 95% limits of agreement for each rater.

Pre-SCI morphometric measurements

In the prone position, cross-sectional areas of the SC and dural sac were relatively consistent among the n = 17 animals, with mean ± SD measurements at the T10 level of 28.91 ± 1.89 mm² and 46.53 ± 4.05 mm², respectively. The ratio between the cross-sectional area of the SC and the dural sac (SC:DS ratio) ranged from 0.55 to 0.69, indicating that the SC occupied 55–69% of the intrathecal space.

Biomechanical injury parameters

All animals (age: 159 ± 31 days; body weight: 25.4 ± 1.9 kg; mean ± SD) received a contusion injury with the 50-g impactor (diameter: 9 mm), dropped from a height of 16.73 cm, followed by 5 min of sustained compression. Peak forces varied between animals and ranged from 2464.0 to 3156.7 kilodynes (Table 1). Upon impact, the impactor tip traveled 4.7 ± 0.3 mm (mean ± SD) from initial contact with the exposed dura at a velocity of 1676.1 ± 30.8 mm/sec, with a maximum displacement of 4.7 ± 0.3 mm.

OPEN IN VIEWER

Table 1.

Measures of Body Weight at Surgery and Biomechanical Parameters of the Contusion Spinal Cord Injury

Animal ID	Cohort (1, 2)	Body weight (kg)	Peak force (kdynes)	Impact velocity (mm/s)	Maximum displacement (mm)
1	1	26.8	2794.1	1667.4	5.0
2	1	27.8	3156.7	1687.3	4.4
3	1	25.7	2768.3	1681.0	4.5
4	2	26.3	2547.5	1646.1	5.0
5	2	21.8	2753.9	1638.4	4.3
6	2	26.7	2905.5	1739.2	4.7
7	2	25.3	2464.0	1640.7	4.5
8	2	23.6	2595.9	1682.3	4.4
9	2	24.2	2630.8	1652.7	4.5
10	2	25.0	2585.8	1665.8	4.5
11	2	23.8	3114.9	1729.3	4.6
12	2	27.4	n.a.	n.a.	n.a.
13	2	23.0	2951.8	1662.1	4.8
14	2	26.1	n.a.	n.a.	n.a.
15	2	28.8	2829.8	1650.0	5.6
16	2	23.6	2751.6	1694.5	4.6
17	2	25.9	2965.7	1704.4	4.7
Mean ± SD		25.4 ± 1.9	2787.8 ± 204.4	1676.1 ± 30.8	4.7 ± 0.3
Range		21.8–28.8	2464.0–3156.7	1638.4–1739.2	4.3–5.6

Animal ID and body weight on the day of SCI surgery are presented for animals in Cohort 1 (n = 3, 6 h post-SCI) and Cohort 2 (n = 14, 7 days post-SCI). Direct measurements of peak force and velocity at impact were recorded by a calibrated load cell located within the impactor tip and further analyzed using LabVIEW software (National Instruments Corporation, Austin, TX). Maximum displacement is the impactor displacement at the initial contact with the exposed dura.

ID, identification; n.a., not available due to a mechanical issue with the impactor; SCI, spinal cord injury; SD, standard deviation.

Gray-value-threshold range determination for IPH segmentation on US

At the impact site, a clearly hyperechoic lesion was observed within both gray and white matter (Supplementary Video S1), which remained visible on US over the following 7 h (Fig. 3). Supplementary Figure S3 presents axial B-mode images compiled for all 12 animals. Examination of the corresponding H&E-stained sections confirmed that areas of increased echogenicity on US within the first 7 h after SCI were associated with regions containing RBCs (see Fig. 4 for an example at 6 h post-SCI).

OPEN IN VIEWER

FIG. 3.

Temporal progression of IPH after SCI on high-frequency US imaging in a porcine model. Representative axial and midsagittal intraoperative B-mode US images from a single animal, acquired at 0.5 h before injury (Pre-SCI), 0.5 h post-injury (HPI), hourly post-injury up to 7 h, and at 168 h post-injury (7 days post-injury). The midsagittal image was identified as the slice where the central canal (CC) was visible. Axial images were captured at the vertical center of the intervertebral disc (indicated by the yellow star). Reconstructed 3D models of the SC (blue) and the IPH (red) are shown for all quantified post-injury time points using the semiautomated segmentation method. During the first 7 h following SCI, IPH was characterized by regions of increased echogenicity throughout the gray matter and white matter, progressing to a more diffuse and less echogenic appearance by 7 days post-injury. The mean rate of 3D IPH surface area expansion, calculated using the semiautomated method, was 8.93 mm²/min during the first 0.5 h post-injury, slowing to 0.54 mm²/min between 0.5 and 7 h post-injury. Imaging parameters: gain: 18 dB; dynamic range: 50 dB; axial slice thickness: 0.203 mm; scanning range: 40 mm. Orientation: D = dorsal; V = ventral; R = right; L = left. IPH, intraparenchymal hemorrhage; SCI, spinal cord injury; US, ultrasound.

OPEN IN VIEWER

FIG. 4.

Histological validation of IPH on US images. Axial intraoperative US images of the SC (upper panel) from a single animal acquired at 6 h post-injury with the corresponding section stained with hematoxylin and eosin (H&E) (lower panel). Linear transformations (i.e., image scaling and rotation) were applied to H&E images (20x objective) to facilitate manual registration of images across modalities. IPH (yellow arrows) appeared as areas of increased echogenicity on US, correlating with the presence of red blood cells (RBCs) on representative H&E-stained sections. US imaging parameters: gain: 22 dB; dynamic range: 25 dB; US slice thickness: 0.203 mm; histology slice thickness: 0.02 mm. Orientation: D = dorsal; V = ventral; R = right; L = left. IPH, intraparenchymal hemorrhage; SC, spinal cord; SCI, spinal cord injury; US, ultrasound.

Although the echogenic appearance of IPH varied over time, the fixed threshold range of 170–248 effectively captured the gray values corresponding to IPH observed during the first 7 h post-SCI (Supplementary Fig. S4 and Fig. 5B). However, IPH within the cord at 7 days post-injury became increasingly difficult to accurately identify using US, and these parameters were found to be unsuitable as they produced false-positive segmentations (Supplementary Fig. S5). As a result, the 7-day post-SCI time point was excluded from subsequent IPH quantification.

OPEN IN VIEWER

FIG. 5.

Determination of the gray-value threshold range for IPH segmentation on US images. (A) To establish the gray-value threshold range of IPH on US, a total of 9 axial intraoperative US images were used (3 images per animal, n = 3) at 6 h post-injury. A region of interest (ROI; red outline) was selected where the hyperechoic signal on US overlapped with RBCs visible on H&E staining. Gray-value histograms were then generated in ImageJ (https://imagej.net/ij/), plotting pixel count (y-axis) against the gray-value range (0–255) on the x-axis. The gray-value range corresponding to each of the 9 ROIs is shown in red. The lowest and highest gray values of the 9 images were used to determine a fixed threshold for IPH segmentation (170–248). (B) The upper panel displays the original US image (left), accompanied by the corresponding H&E-stained section (middle), alongside the resulting IPH segmentation achieved using the semiautomated approach with a threshold range of 170–248 (right). The lower panel presents a magnified view of ROIs 1 and 2, illustrating that areas of increased echogenicity on US—indicative of the presence of RBCs—were successfully segmented as IPH (red). Imaging parameters: gain: 22 dB; dynamic range: 25 dB; axial slice thickness: 0.203 mm. Orientation: D = dorsal; V = ventral; R = right; L = left. H&E, hematoxylin and eosin; IPH, intraparenchymal hemorrhage; RBCs, red blood cells; SC, spinal cord; SCI, spinal cord injury; US, ultrasound.

IPH quantification within the first 7 h post-SCI

Using the above-established threshold value, we then applied our semiautomated method to measure the volumetric dimensions of the SC and IPH in 12 animals during the hyperacute phase following SCI (0.5–7 h) (Fig. 6).

OPEN IN VIEWER

FIG. 6.

Quantification of SC swelling and IPH progression over 7 h using high-frequency US in a porcine model of SCI. SC swelling and IPH progression were monitored and quantified using intraoperative 3D B-mode US images obtained at multiple time points: 0.5 h before injury (pre-SCI), after SC decompression (0.5 h post-injury), and at 2, 4, and 7 h post-injury. Utilizing our semiautomated segmentation and quantification approach, the following outcome measures were obtained: (A) SC volume (mm³), (B) IPH volume (mm³), (C) IPH volume as a percentage of SC volume over a 40-mm segment, (D) 3D IPH SA (mm²), and (E) IPH length (mm). Data are presented for individual animals (n = 12) with solid gray lines, and red symbols and lines indicate the median with interquartile range (IQR). (F) The 2D IPH area (mm²) per axial US slice plotted against the distance from the epicenter (0 mm) at different post-injury time points: 0.5 h (pink), 2 h (red), 4 h (orange), and 7 h (brown). Imaging parameters: gain: 22 dB; dynamic range: 25 dB; slice thickness: 0.203 mm; scanning range: 40 mm. Only statistically significant p values (p < 0.05) are displayed. ***p < 0.001. IPH, intraparenchymal hemorrhage; SC, spinal cord; SCI, spinal cord injury; US, ultrasound.

Most animals exhibited SC deformation within 0.5 h, followed by swelling that increased 1.14-fold at 7 h post-injury compared with pre-injury measurements (p < 0.001) (Fig. 6A). IPH increased rapidly in the first 30 min post-SCI at approximately 1.1 mm³/min between SCI and 0.5 h (Fig. 7B), reaching about 34.0 mm³ (interquartile range [IQR]: 8.9), which is 3.8% (IQR: 0.9) of the volume of a 40-mm SC segment (Fig. 6B, C).

OPEN IN VIEWER

FIG. 7.

Calculated rate of SC and IPH expansion during the first 7 h post-SCI. Utilizing our semiautomated segmentation and quantification approach, we calculated the rate of expansion for: (A) SC volume (mm³), (B) IPH volume (mm³), (C) IPH volume as a percentage of SC volume over a 40-mm segment, (D) 3D IPH SA (mm²), and (E) IPH length (mm). Expansion rate was determined by dividing the change in volume, percentage, area, or length between the initial and follow-up scans by the time interval between the two imaging sessions. Data are presented as box-and-whisker plots, showing the mean, 25th–75th percentiles, minimum, and maximum, with individual data points overlaid. Only statistically significant p values (p < 0.05) are displayed. ***p < 0.001. IPH, intraparenchymal hemorrhage; SA, surface area; SC, spinal cord; SCI, spinal cord injury; US, ultrasound.

SC volume at 0.5 h post-SCI was not correlated with IPH volume at 0.5 h post-SCI (rho = −0.04, p = 0.904), indicating that swelling at 0.5 h was not proportional to IPH volume at the same time (Supplementary Fig. S6). However, SC volume at 0.5 h post-SCI was significantly correlated with the change in IPH volume between 0.5 and 2 h post-SCI (rho = 0.62, p = 0.037), suggesting that swelling at 0.5 h was associated with IPH expansion within the first 2 h.

By 7 h, IPH volume reached 48.9 mm³ (IQR: 15.0), representing a 1.44-fold increase from 30 min (p < 0.001), along with a concurrent increase in 3D IPH SA (Fig. 6D). The expansion rate decreased to 0.036 mm³/min between 2 and 7 h, with localized growth at the impact site rather than elongation along the rostral–caudal axis (Fig. 6E, F).

Relationship between morphometric, biomechanical, and IPH outcomes

Next, we performed Spearman’s rank correlations between the pre-SCI morphometric, biomechanical, and IPH outcomes, focusing on the early (0.5 h post-SCI) and late changes (7 h post-SCI) (Supplementary Fig. S7). At 7 h post-SCI, significant moderate correlations were observed between peak force and absolute IPH volume (rho = 0.61, p = 0.033), percentage IPH (rho = 0.66, p = 0.021), and 3D IPH SA (rho = 0.58, p = 0.044).

Performance testing: varying overall receiver gain and dynamic range settings

Subsequently, we examined how variations in US image contrast affect the visualization and quantification of IPH characteristics. We specifically compared a high-contrast scenario characterized by a gain setting of 22 dB and a dynamic range of 25 dB to a lower-contrast scenario, which featured a reduced gain of 18 dB and a widened dynamic range of 50 dB (Supplementary Fig. S8).

Upon analysis, significant discrepancies were observed between acquisition settings in both the measured IPH volume (p = 0.008) and SA (p = 0.008). Notably, the higher contrast settings yielded greater average measurements for both parameters.

Performance testing: Manual versus semiautomated IPH quantification

Exploratory analyses of pooled measurements across animals showed that manual and semiautomated 2D IPH area measurements were significantly correlated, with Spearman’s rho ranging from 0.87 to 0.95 (Fig. 8A). Exploratory Bland–Altman analyses indicated mean biases between manual and semiautomated IPH area measurements as follows: Rater 1, 1.11 mm² (95% limits of agreement: −0.58 to 2.81 mm²); Rater 2, 1.80 mm² (0.29–3.30 mm²); and Rater 3, 1.04 mm² (−1.37 to 3.45 mm²) (Fig. 8B).

OPEN IN VIEWER

FIG. 8.

Comparison of manual and semiautomated IPH area measurements. Exploratory analyses indicated that IPH area (mm²) measurements were consistently larger with manual segmentation than semiautomated segmentation. (A) Scatter plots comparing IPH area (mm²) measurements obtained using manual and semiautomated segmentation methods over a scanning range of 40 mm. Spearman’s rho and p values are provided alongside regression lines for each rater individually. (B) Bland-Altman plots showing agreement between manual and semiautomated IPH area measurements. The mean difference (manual minus semiautomated IPH area) is represented by the thick black horizontal line, and the 95% limits of agreement (±1.96 standard deviation [SD]) are represented by the black dotted lines for each rater. IPH, intraparenchymal hemorrhage; SC, spinal cord; SCI, spinal cord injury; US, ultrasound.

Inter-rater reliability was assessed across 15 axial US images (from n = 3 animals, 5 images each). Pooled ICCs indicated excellent agreement for both manual (ICC = 0.967, 95% confidence interval [CI]: 0.917–0.988) and semiautomated segmentation (ICC = 0.999, 95% CI: 0.997–1.000) (Table 2). Mixed-effects ICCs accounting for repeated measures confirmed excellent inter-rater reliability (manual: 0.907; semiautomated: 0.996) and intra-rater reliability (manual: 0.922; semiautomated: 0.996).

OPEN IN VIEWER

Table 2.

Comparison of Inter-Rater and Intra-Rater Reliability for Intraparenchymal Hemorrhage Area Measures: Manual Versus Semiautomated

Area measures	Manual		Semiautomated
	ICC (95% CI)	p Value	ICC (95% CI)	p Value
Inter-rater reliability
SC area (mm2)	0.853 (0.324–0.958)	<0.001	0.823 (0.340–0.946)	<0.001
IPH area (mm2)	0.967 (0.917–0.988)	<0.001	0.999 (0.997–1.000)	<0.001
IPH (% SC area)	0.969 (0.926–0.989)	<0.001	0.998 (0.994–0.999)	<0.001
Intra-rater reliability
SC area (mm2)
Rater 1	0.851 (0.622–0.947)	<0.001	0.812 (0.211–0.946)	<0.001
Rater 2	0.887 (0.703–0.960)	<0.001	0.718 (0.160–0.908)	<0.001
Rater 3	0.920 (0.777–0.972)	<0.001	0.834 (0.095–0.957)	<0.001
IPH area (mm2)
Rater 1	0.894 (0.276–0.973)	<0.001	0.996 (0.989–0.999)	<0.001
Rater 2	0.983 (0.950–0.994)	<0.001	0.998 (0.995–0.999)	<0.001
Rater 3	0.945 (0.828–0.982)	<0.001	0.997 (0.975–0.999)	<0.001
IPH (% SC area)
Rater 1	0.889 (0.251–0.972)	<0.001	0.996 (0.987–0.999)	<0.001
Rater 2	0.984 (0.939–0.995)	<0.001	0.996 (0.989–0.999)	<0.001
Rater 3	0.948 (0.840–0.983)	<0.001	0.995 (0.896–0.999)	<0.001

Three raters executed the manual and semiautomated segmentation protocols on a subset of 15 axial images from 3 animals (5 images per animal) to obtain inter-rater reliability values and intra-rater reliability values across 2 sessions (Session 1 and Session 2), separated by at least 1 month. Images acquired at 4 h post-SCI were used for analysis.

CI, confidence interval; ICC, intraclass coefficient; IPH, intraparenchymal hemorrhage; SC, spinal cord.

Regarding inter-rater agreement, semiautomated segmentation exhibited a higher degree of consensus, with 96% of the total IPH area identified by all three raters compared to 55% with manual segmentation (Table 3). Additionally, semiautomated segmentation yielded higher DSC values than manual segmentation (Table 4). Most discrepancies between the manual and semiautomated methods occurred at the outer borders of the IPH, particularly within the white matter, where the IPH was more diffuse or had less distinct boundaries. These areas are highlighted in Figure 9 as pink and purple regions, indicating assessments by one-third or two-thirds of the raters. Notably, this pattern was predominantly observed in manual segmentation.

OPEN IN VIEWER

Table 3.

Percentage of Total Intraparenchymal Hemorrhage Area Identified by All Three Raters Using Both Manual and Semiautomated Segmentation Methods

	Percentage of total IPH area identified (all three raters)					Mean percentage of total IPH area identified (all three raters) ± SD
Animal and IPH analysis method	Slice 1	Slice 2	Slice 3	Slice 4	Slice 5	Slices 1–5
Animal 1
Manual	51	67	57	56	46	55 ± 8
Semiautomated	95	98	99	97	91	96 ± 3
Animal 2
Manual	39	49	65	50	51	51 ± 9
Semiautomated	82	89	98	92	85	89 ± 6
Animal 3
Manual	43	65	65	57	44	55 ± 11
Semiautomated	75	88	99	92	89	89 ± 9

Each rater independently segmented a subset of 15 axial images from 3 animals (5 images per animal). For each selected image, the table reports the percentage of the total IPH area segmented by all 3 raters (3 out of 3) across the 2 segmentation techniques.

IPH, intraparenchymal hemorrhage; SD, standard deviation.

OPEN IN VIEWER

Table 4.

Dice Similarity Coefficient Scores for Intraparenchymal Hemorrhage Area Measurements Obtained Through Manual and Semiautomated Segmentation

	DSC					Mean DSC ± SD
Animal, IPH analysis method, and rater comparison	Slice 1	Slice 2	Slice 3	Slice 4	Slice 5	Slices 1–5
Manual
Animal 1
Rater 1 vs. 2	0.74	0.85	0.82	0.80	0.74	0.79 ± 0.05
Rater 1 vs. 3	0.74	0.87	0.79	0.81	0.74	0.79 ± 0.05
Rater 2 vs. 3	0.85	0.88	0.84	0.82	0.76	0.83 ± 0.04
Animal 2
Rater 1 vs. 2	0.75	0.73	0.85	0.77	0.73	0.76 ± 0.05
Rater 1 vs. 3	0.69	0.81	0.89	0.80	0.85	0.81 ± 0.08
Rater 2 vs. 3	0.66	0.77	0.84	0.76	0.77	0.76 ± 0.06
Animal 3
Rater 1 vs. 2	0.72	0.85	0.86	0.85	0.76	0.81 ± 0.06
Rater 1 vs. 3	0.80	0.93	0.86	0.81	0.71	0.82 ± 0.08
Rater 2 vs. 3	0.91	0.90	0.86	0.80	0.74	0.84 ± 0.07
Semiautomated
Animal 1
Rater 1 vs. 2	0.98	1.00	0.99	1.00	0.98	0.99 ± 0.01
Rater 1 vs. 3	0.98	0.99	1.00	0.98	0.97	0.98 ± 0.01
Rater 2 vs. 3	0.99	0.99	0.99	0.99	0.96	0.98 ± 0.02
Animal 2
Rater 1 vs. 2	0.91	0.95	0.99	0.96	0.92	0.95 ± 0.03
Rater 1 vs. 3	0.92	0.95	0.99	0.99	0.99	0.97 ± 0.03
Rater 2 vs. 3	0.97	0.98	1.00	0.97	0.93	0.97 ± 0.03
Animal 3
Rater 1 vs. 2	0.86	0.94	0.99	0.96	0.98	0.95 ± 0.05
Rater 1 vs. 3	0.99	0.97	1.00	0.97	0.94	0.98 ± 0.02
Rater 2 vs. 3	0.87	0.96	0.99	0.98	0.96	0.95 ± 0.05

Three raters executed the manual and semiautomated segmentation protocols on a subset of 15 axial images from 3 animals (5 images per animal). DSC values were calculated for each pair of raters by doubling the area of intersection of their IPH segmentations, divided by the total IPH area. For each selected image, the individual DSC values are reported, along with the mean DSC ± SD, averaged across Slices 1–5.

DSC, dice similarity coefficient; IPH, intraparenchymal hemorrhage; SD, standard deviation.

OPEN IN VIEWER

FIG. 9.

Comparison of manual and semiautomated IPH segmentation maps for area agreement. Semiautomated IPH segmentation showed higher inter-rater agreement among raters compared to manual segmentation. (A) Original intraoperative axial B-mode US images acquired 4 h post-injury from 1 of the 3 datasets used to evaluate the performance of the semiautomated method. Orientation: D = dorsal; V = ventral; R = right; L = left. (B, C) Percentage of total IPH area identified by all 3 raters (yellow), 2 out of 3 raters (pink), and 1 out of 3 raters (purple) using manual segmentation and semiautomated segmentation for each of the 5 slices. Imaging parameters: gain: 22 dB; dynamic range: 25 dB; axial slice thickness: 0.203 mm. IPH, intraparenchymal hemorrhage; US, ultrasound.

Performance testing: IPH volumetric quantification using the semiautomated IPH method

The semiautomated approach showed excellent inter- and intra-rater ICC values for IPH and SC volume measurements, indicating high reliability (ICC range: 0.961–0.980) (Table 5). Raters with different US experience achieved consensus, with agreement ranging from 91.1% to 91.4% (Fig. 10) and DSC scores greater than 0.9 (Table 6), confirming strong agreement in IPH segmentation.

OPEN IN VIEWER

Table 5.

Comparison of Inter-Rater and Intra-Rater Reliability for Intraparenchymal Hemorrhage Volumetric Measures Using Semiautomated Segmentation

Volumetric measures	Semiautomated
	ICC (95% CI)	p Value
Inter-rater reliability
SC volume (mm3)	0.976 (0.556–0.999)	<0.001
IPH volume (mm3)	0.980 (0.665–0.999)	<0.001
IPH (% SC volume)	0.961 (0.443–0.999)	<0.001
Intra-rater reliability
SC volume (mm3)
Rater 1	0.973 (0.360–0.999)	0.006
Rater 2	0.948 (0.015–0.999)	0.006
Rater 3	0.992 (0.850–1.000)	0.005
IPH volume (mm3)
Rater 1	0.998 (0.847–1.000)	<0.001
Rater 2	0.934 (0.180–0.998)	0.020
Rater 3	0.972 (0.040–0.999)	<0.001
IPH (% SC volume)
Rater 1	0.985 (0.384–1.000)	0.002
Rater 2	0.874 (−0.066–0.996)	0.027
Rater 3	0.956 (0.020–0.999)	<0.001

Three raters executed the manual and semiautomated segmentation protocols on the full dataset (197 images per animal, n = 3) to obtain inter- and intra-rater reliability values across Sessions 1 and 2. Images acquired at 4 h post-SCI were used for analysis.

CI, confidence interval; ICC, intraclass correlation coefficient; IPH, intraparenchymal hemorrhage; SC, spinal cord; SCI, spinal cord injury.

OPEN IN VIEWER

FIG. 10.

Semiautomated IPH segmentation maps for volume agreement. IPH volume segmentation maps are shown for (A) Session 1 and (B) Session 2, separated by at least 1 month. The percentage of total IPH volume over a 40-mm cord segment identified by all 3 raters (yellow), 2 out of 3 raters (pink), and 1 out of 3 raters (purple) using semiautomated segmentation is presented for the entire dataset (197 slices per animal, n = 3). 3D SC mask = blue. Orientation: D = dorsal; V = ventral; R = rostral; C = caudal. IPH, intraparenchymal hemorrhage.

OPEN IN VIEWER

Table 6.

Dice Similarity Coefficient for Intraparenchymal Hemorrhage Volume Measures Using Semiautomated Segmentation

Inter-rater agreement	DSC	Intra-rater agreement	DSC
Animal 1		Animal 1
Rater 1 vs. 2	0.98	Rater 1 vs. 1	0.99
Rater 1 vs. 3	0.98	Rater 2 vs. 2	0.98
Rater 2 vs. 3	0.98	Rater 3 vs. 3	0.97
Animal 2		Animal 2
Rater 1 vs. 2	0.96	Rater 1 vs. 1	0.98
Rater 1 vs. 3	0.97	Rater 2 vs. 2	0.97
Rater 2 vs. 3	0.96	Rater 3 vs. 3	0.97
Animal 3		Animal 3
Rater 1 vs. 2	0.95	Rater 1 vs. 1	0.98
Rater 1 vs. 3	0.97	Rater 2 vs. 2	0.94
Rater 2 vs. 3	0.96	Rater 3 vs. 3	0.96

Three raters executed the semiautomated segmentation protocol (197 images per animal, n = 3). The DSC was calculated for each pair of raters as twice the volume overlap between the IPH segmentations, divided by the total IPH segmentation volume.

DSC, dice similarity coefficient; IPH, intraparenchymal hemorrhage.