Automated detection and analysis of subdural hematomas using a machine learning algorithm

Marco Colasurdo Department of Radiology, Division of Neuroradiology, The University of Texas Medical Branch, Galveston, Texas;

Search for other papers by Marco Colasurdo in
Current site
Google Scholar
PubMed
Close
 MD
,
Nir Leibushor Viz.ai Inc., San Francisco, California;

Search for other papers by Nir Leibushor in
Current site
Google Scholar
PubMed
Close
 BSc
,
Ariadna Robledo Departments of Neurosurgery and

Search for other papers by Ariadna Robledo in
Current site
Google Scholar
PubMed
Close
 MD
,
Viren Vasandani Departments of Neurosurgery and

Search for other papers by Viren Vasandani in
Current site
Google Scholar
PubMed
Close
 MD
,
Zean Aaron Luna Departments of Neurosurgery and

Search for other papers by Zean Aaron Luna in
Current site
Google Scholar
PubMed
Close
 BSc
,
Abhijit S. Rao Departments of Neurosurgery and

Search for other papers by Abhijit S. Rao in
Current site
Google Scholar
PubMed
Close
 BSA
,
Roberto Garcia Departments of Neurosurgery and

Search for other papers by Roberto Garcia in
Current site
Google Scholar
PubMed
Close
 MD
,
Visish M. Srinivasan Department of Neurosurgery, Barrow Neurological Institute, St. Joseph’s Hospital and Medical Center, Phoenix, Arizona; and

Search for other papers by Visish M. Srinivasan in
Current site
Google Scholar
PubMed
Close
 MD
,
Sunil A. Sheth Department of Neurology, McGovern Medical School, University of Texas Health Science Center, Houston, Texas

Search for other papers by Sunil A. Sheth in
Current site
Google Scholar
PubMed
Close
 MD
,
Naama Avni Viz.ai Inc., San Francisco, California;

Search for other papers by Naama Avni in
Current site
Google Scholar
PubMed
Close
 BSc
,
Moleen Madziva Viz.ai Inc., San Francisco, California;

Search for other papers by Moleen Madziva in
Current site
Google Scholar
PubMed
Close
 PhD
,
Mor Berejick Viz.ai Inc., San Francisco, California;

Search for other papers by Mor Berejick in
Current site
Google Scholar
PubMed
Close
 BSc
,
Goni Sirota Viz.ai Inc., San Francisco, California;

Search for other papers by Goni Sirota in
Current site
Google Scholar
PubMed
Close
 BSc
,
Aielet Efrati Viz.ai Inc., San Francisco, California;

Search for other papers by Aielet Efrati in
Current site
Google Scholar
PubMed
Close
 PhD
,
Avraham Meisel Viz.ai Inc., San Francisco, California;

Search for other papers by Avraham Meisel in
Current site
Google Scholar
PubMed
Close
 HSD
,
Hashem Shaltoni Neurology, The University of Texas Medical Branch, Galveston, Texas;

Search for other papers by Hashem Shaltoni in
Current site
Google Scholar
PubMed
Close
 MD
, and
Peter Kan Departments of Neurosurgery and

Search for other papers by Peter Kan in
Current site
Google Scholar
PubMed
Close
 MD, MPH
Free access

OBJECTIVE

Machine learning algorithms have shown groundbreaking results in neuroimaging. Herein, the authors evaluate the performance of a newly developed convolutional neural network (CNN) to detect and quantify the thickness, volume, and midline shift (MLS) of subdural hematoma (SDH) from noncontrast head CT (NCHCT).

METHODS

NCHCT studies performed for the evaluation of head trauma in consecutive patients between July 2018 and April 2021 at a single institution were retrospectively identified. Ground truth determination of SDH, thickness, and MLS was established by the neuroradiology report. The primary outcome was performance of the CNN in detecting SDH in an external validation set, as measured using area under the receiver operating characteristic curve analysis. Secondary outcomes included accuracy for thickness, volume, and MLS.

RESULTS

Among 263 cases with valid NCHCT according to the study criteria, 135 patients (51%) were male, the mean (± standard deviation) age was 61 ± 23 years, and 70 patients were diagnosed with SDH on neuroradiologist evaluation. The median SDH thickness was 11 mm (IQR 6 mm), and 16 patients had a median MLS of 5 mm (IQR 2.25 mm). In the independent data set, the CNN performed well, with sensitivity of 91.4% (95% CI 82.3%–96.8%), specificity of 96.4% (95% CI 92.7%–98.5%), and accuracy of 95.1% (95% CI 91.7%–97.3%); sensitivity for the subgroup with an SDH thickness above 10 mm was 100%. The maximum thickness mean absolute error was 2.75 mm (95% CI 2.14–3.37 mm), whereas the MLS mean absolute error was 0.93 mm (95% CI 0.55–1.31 mm). The Pearson correlation coefficient computed to determine agreement between automated and manual segmentation measurements was 0.97 (95% CI 0.96–0.98).

CONCLUSIONS

The described Viz.ai SDH CNN performed exceptionally well at identifying and quantifying key features of SDHs in an independent validation imaging data set.

ABBREVIATIONS

AI = artificial intelligence; CNN = convolutional neural network; LVO = large vessel occlusion; MLS = midline shift; NCHCT = noncontrast head CT; NPV = negative predictive value; PPV = positive predictive value; SDH = subdural hematoma.

OBJECTIVE

Machine learning algorithms have shown groundbreaking results in neuroimaging. Herein, the authors evaluate the performance of a newly developed convolutional neural network (CNN) to detect and quantify the thickness, volume, and midline shift (MLS) of subdural hematoma (SDH) from noncontrast head CT (NCHCT).

METHODS

NCHCT studies performed for the evaluation of head trauma in consecutive patients between July 2018 and April 2021 at a single institution were retrospectively identified. Ground truth determination of SDH, thickness, and MLS was established by the neuroradiology report. The primary outcome was performance of the CNN in detecting SDH in an external validation set, as measured using area under the receiver operating characteristic curve analysis. Secondary outcomes included accuracy for thickness, volume, and MLS.

RESULTS

Among 263 cases with valid NCHCT according to the study criteria, 135 patients (51%) were male, the mean (± standard deviation) age was 61 ± 23 years, and 70 patients were diagnosed with SDH on neuroradiologist evaluation. The median SDH thickness was 11 mm (IQR 6 mm), and 16 patients had a median MLS of 5 mm (IQR 2.25 mm). In the independent data set, the CNN performed well, with sensitivity of 91.4% (95% CI 82.3%–96.8%), specificity of 96.4% (95% CI 92.7%–98.5%), and accuracy of 95.1% (95% CI 91.7%–97.3%); sensitivity for the subgroup with an SDH thickness above 10 mm was 100%. The maximum thickness mean absolute error was 2.75 mm (95% CI 2.14–3.37 mm), whereas the MLS mean absolute error was 0.93 mm (95% CI 0.55–1.31 mm). The Pearson correlation coefficient computed to determine agreement between automated and manual segmentation measurements was 0.97 (95% CI 0.96–0.98).

CONCLUSIONS

The described Viz.ai SDH CNN performed exceptionally well at identifying and quantifying key features of SDHs in an independent validation imaging data set.

In Brief

The authors studied the performance of a newly developed convolutional neural network (CNN) for automated detection and characterization of subdural hematoma (SDH). The Viz.ai SDH CNN performed exceptionally well at identifying and quantifying key features of SDH in an independent validation imaging data set. The findings will lay the foundation for future development of the CNN including more complex assessments with additional features that go far beyond the simple image analysis task.

Subdural hematoma (SDH) is a common condition encountered in clinical practice and is described as a pathological collection that accumulates in the potential space between the dura mater and arachnoid and has a classic crescentic morphology when localized at the cerebral convexity. It is classically categorized as acute, subacute, chronic, or acute on chronic based on the time of presentation, clinical characteristics, and imaging features. In addition, classifications with detailed subgroups have also been proposed such as the one from Nakaguchi et al.,1 which is specifically designed for predicting the postoperative risk of chronic SDH.

Chronic SDHs are typically seen in elderly patients, with an insidious onset and progression, and their pathophysiology is still far from being fully understood. Their overall reported incidence ranges from 1.72% to 20.6% with increasing rates given the aging population and the rising prevalence of anticoagulation/antiplatelet medication use.2 Morbidity and mortality rates vary greatly with reported rates ranging from 0% to 25% and from 0% to 32%, respectively.36 Acute SDHs, on the other hand, are commonly associated with traumatic brain injury,7 which is responsible for half of all deaths of head injury patients regardless of age.8 Despite improvements in management and therapy, mortality in patients with acute SDH can still be as high as 50% to 70%.911

Triaging SDH relies heavily on imaging, and noncontrast head CT (NCHCT) evaluation is currently the standard of care that allows initial diagnosis of SDH. NCHCT findings can further predict raised intracranial pressure through the size of the hematoma, effacement of the cisterns, or midline shift (MLS), allowing one to decide between the need for surgical evacuation or conservative therapy. However, this critical process is subjective given differences in reader experience and can be limited by reading times during on-call hours, potentially leading to delayed diagnosis or incorrect interpretations.

Artificial intelligence (AI) and deep learning can facilitate the approach to these situations. Numerous AI applications are in development or in use in different fields of medicine, for example, in neuroscience for acute stroke for the detection of large vessel occlusion (LVO) and management of endovascular thrombectomy.12,13 With this in mind, we note that automated systems for the detection and analysis of SDH can enable earlier and more accurate diagnosis of pathological findings, alerting both the radiologist and the clinician to the need for increasing the prioritization of the read,14 earlier evaluation of the patient, and potentially earlier intervention. The use of automated measurement and volume segmentation of SDH can also be foreseen for clinical trial screening and analysis of large data sets outside the clinical environment.

Viz.ai has developed a deep convolutional neural network (CNN) designed to detect and analyze SDH in patients of all ages on NCHCT. In this paper, we outline the CNN and assess its performance in detecting the presence or absence of SDH as well as evaluating key characteristics of SDH.

Methods

This study was compliant with the Health Insurance Portability and Accountability Act and was approved by the institutional review board. The requirement for patient consent was waived because of the investigation’s retrospective nature and the de-identification of patient data. Nonemployee and nonconsultant authors (M.C., A.R., V.V., Z.A.L., A.S.R., V.M.S., S.A.S., H.S., P.K.) had complete control over data, information, and analysis.

Cohort Characteristics

We retrospectively reviewed NCHCT studies from 340 consecutive cases with a history of acute or chronic trauma from July 2018 to April 2021 at the same institution (University of Texas Medical Branch at Galveston).

Imaging Settings

The images used to develop and test the algorithm had been acquired in the axial plane on six scanners by different manufacturers. The following criteria were inspected to validate technical and clinical adequacy: axial monochrome NCHCT scan, number of slices between 18 and 125, slice thickness between 2.5 and 5 mm, scan dimension 512 × 512 pixels, whole-head volume included without excessive motion artifacts, and recent postoperative changes. Images were stored on a PACS.

Pipeline for Analysis

The training data set utilized to develop the CNN included 1429 nonconsecutive NCHCT scans that had been obtained from October 2018 through June 2021 in patients with and without intracranial hemorrhage. In particular, 251 patients had uni- or bilateral SDH, 762 had different kinds of intracranial hemorrhage (intraparenchymal hematoma, intraventricular hemorrhage, subarachnoid hemorrhage, or epidural hematoma), and 416 did not have any intracranial hemorrhage. Forty-four percent of the training data set subjects were female, and the mean (± standard deviation [SD]) patient age was 62 ± 16 years. Forty different types of scanners had been used with a slice thickness varying from 2.5 to 5 mm.

The testing data set utilized 340 unique NCHCT scans, 263 of which were deemed valid based on our imaging criteria. Seventy scans had an SDH, and 193 scans were considered negative based on the neuroradiologist read (Fig. 1).

FIG. 1.
FIG. 1.

Diagram illustrating the included and excluded cases in the analyzed data sets.

Ground truth on the training data set was established based on the final reports provided by dedicated, experienced neuroradiologists compared with the read provided by a specialized clinical annotator. A dedicated neuroradiologist with 5 years of experience hand-contoured each hematoma separately on NCHCT using the open-source software ITK-SNAP with procedures described in prior publications (http://www.itksnap.org/pmwiki/pmwiki.php).15

Ground truth on the testing data set was established by a specialized clinical annotator. It was then compared with the final read provided by 5 dedicated neuroradiologists with 3, 5, 6, 7, and 10+ years of dedicated neuroradiology experience. In 6 cases the clinical annotator’s read differed from the read provided by the neuroradiologist, and a third independent read by a senior neurosurgeon with 10+ years of experience and blinded to the prior analysis was used as a discriminating factor.

CNN Description

Viz.ai SDH is based on a deep CNN that receives an NCHCT scan as a 3D array and outputs a probability mask for the presence of a subdural hemorrhage in each voxel. The probabilities are thresholded to obtain a binary mask. Then, the SDH volume (in ml) is calculated based on the binary mask, using the pixel spacing and slice thickness.

The SDH maximum thickness is calculated as the maximal vertical length between the bleed segmentation and bone pixels. The thickness calculation is executed on slices above the temporal bones and up to two slices above the lateral ventricles. The upper limit is taken to exclude miscalculation on high skull curvature. The lower limit avoids measurement on tentorial bleeding and is performed by excluding slices lower than one below the ventricles. Bleed segments in the falx are excluded from this calculation. Ventricle identification is done using a dedicated segmentation model.

The MLS measurement is calculated by segmenting several anatomical midline structures: the lateral ventricles, the septum pellucidum, the cavum septum pellucidum if present, the third ventricle, the posterior and anterior sinuses, and the skull crests. The relevant slice featuring the foramen of Monro is identified as the highest slice that contains segmentation of the third ventricle. The midline in its normal position is drawn from the posterior crest/sinus to the anterior crest/sinus, and the shift is measured as the vertical distance from the septum pellucidum middle to this theoretical line.

Statistical Analysis

Statistical analyses were performed using descriptive data, including ranges, means, medians, and standard deviations for continuous variables and frequencies and percentages for categorical variables. System performance methods were examined using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and total accuracy. Confidence intervals were calculated using the Wilson procedure with correction for continuity.16 Receiver operating characteristic curve analysis was used to determine the characteristics and performance of the automated SDH volume measurement. A Pearson correlation coefficient and Bland-Altman plot were used to determine agreement between manual and automated segmentation measurements.

Results

Among 263 cases included in the study, the mean (± SD) patient age was 61 ± 23 years, and 135 patients (51%) were male. There were 35 chronic SDHs (50%), while the remaining 50% comprised acute (6 [8.6%]), subacute (3 [4.3%]), and acute on chronic (26 [37.1%]) SDHs. There were no missing data in the analyzed cohort (Fig. 1).

Detection and Thickness Measurement

System performance for the entire group was measured, showing sensitivity of 91.4% (95% CI 82.3%–96.8% with 64/70 cases), specificity of 96.4% (95% CI 92.7%–98.5% with 186/193 cases), NPV of 96.9% (95% CI 93.3%–98.8%), PPV of 90.1% (95% CI 80.7%–95.9% with 64/70 cases), and accuracy of 95.1% (95% CI 91.7%–97.3% with 250/263 cases). The maximum SDH thickness mean absolute error was 2.75 mm (95% CI 2.14–3.37 mm).

Moreover, in the subgroup with an SDH thickness above 10 mm, sensitivity was 100% (95% CI 90.0%–100.0% with 44/44 cases); in the group with an SDH thickness above 5 mm, sensitivity was 96.7% (95% CI 87.5%–99.4% with 58/60 cases). System performance for the acute, subacute, or acute on chronic subgroup had a sensitivity of 94.1% (95% CI 78.9%–99.0%) and for the chronic subgroup had a sensitivity of 88.6% (95% CI 72.3%–96.2%).

MLS and Volume Measurement

The newly developed CNN identified rightward or leftward MLS in 100% of cases with a mean absolute error of 0.93 mm (95% CI 0.55–1.31 mm).

Receiver operating characteristic analysis demonstrated an area under the curve of 96.9% for SDH volume (Fig. 2).

FIG. 2.
FIG. 2.

Receiver operating characteristic curve analysis plot based on SDH volume showing an area under the curve (AUC) above 95%. Figure is available in color online only.

A Pearson correlation coefficient was computed to assess the linear relationship between automated and manual segmentation measurements. There was a positive correlation between the two variables (R2 = 0.9485, Pearson coefficient = 0.97, 95% CI 0.96–0.98; Fig. 3A). Agreement between automated and manual hand-contoured segmentations is also demonstrated in the Bland-Altman plot (Fig. 3B).

FIG. 3.
FIG. 3.

Pearson correlation coefficient (A) and Bland-Altman plot (B) comparing manual and automated volume segmentations. Figure is available in color online only.

Examples of correct flagging, segmentation, and thickness and volume measurements for 2 patients are provided in Fig. 4. A challenging segmentation is shown in Fig. 4B, a case in which a patient with a left SDH also presented with an intraparenchymal hematoma and subarachnoid hemorrhage. Additional examples of maximum thickness measurement and the automated MLS feature are demonstrated in Fig. 5A and B. An example of infratentorial segmentation is also provided in Fig. 5C. Examples of incorrect flagging are provided in Fig. 5D–F.

FIG. 4.
FIG. 4.

A: Example of algorithm output in a case with bilateral SDHs, with volume and maximum thickness measurements. B: Additional example of segmentation and maximum thickness measurement in a patient with a left-sided acute SDH, a left frontal intraparenchymal hemorrhage, and ipsilateral frontoparietal subarachnoid hemorrhage. Figure is available in color online only.

FIG. 5.
FIG. 5.

MLS deviation feature in a patient with bilateral SDHs (A) and right-sided SDH (B). Additional example of segmentation and maximum thickness measurement in a patient with an SDH along the falx cerebri (C). Common diagnostic pitfalls flagged as an SDH are enlarged pericephalic subarachnoid spaces (D). Hyperdense asymmetrical pericerebral subarachnoid spaces (E) and hyperdense tentorial leaflets (F) can be additional examples of false-positive flagging. Figure is available in color online only.

Discussion

In this study, we found the newly developed Viz.ai SDH CNN to be accurate in the detection of SDH. Its performance in detecting the presence or absence of SDH was extremely accurate17 with an area under the curve for hematoma volume above 0.9 in an independent validation imaging data set. It also performed similarly well in secondary outcomes such as MLS evaluation, in which it correctly identified rightward or leftward MLS in 100% of cases with a mean absolute error of 0.93 mm.

The CNN encountered 7 false-positive and 6 false-negative results. The most commonly encountered false-positive pitfalls are misinterpretation of increased pericerebral subarachnoid spaces and large encephalomalacic postsurgical changes erroneously considered to be chronic SDH. In those cases in which it is difficult to ascertain if there is a chronic SDH or an increase of the pericerebral subarachnoid spaces related to brain atrophy, even dedicated neuroradiologists will sometimes resort to the use of prior available imaging or request further follow-up scans. Additionally, a rarer instance of a false-positive example is represented by patients with dense tentorial leaflets, often seen in younger or dehydrated individuals and could be misinterpreted as an acute SDH, a common pitfall even for experienced general radiologists. The 6 false-negative cases were all nonacute SDH (6 chronic and 1 subacute) and had a maximum thickness of less than 10 mm.

When considering only clinically significant SDH with a maximum thickness above 10 mm and 5 mm, the algorithm respectively demonstrated 100% sensitivity with 35/35 cases correctly identified and 96.7% sensitivity with 58/60 cases correctly identified.

NCHCT is the modality of choice for the diagnosis of SDH. The presence of hemorrhage is what discriminates between the need for admission and further evaluation or discharge of patients. Data derived from NCHCT reads are type (acute or chronic), location, maximum thickness, gross volume, and the presence and quantification of MLS. These factors are crucial for clinical or interventional management, and the lack of a readily available trained and experienced clinician could negatively impact management and consequently clinical outcome.

MLS measurement is one of the most important features used to evaluate the severity of brain compression and is routinely performed by clinicians and radiologists at the level of the foramen of Monro, on the basis of axial images. The CNN-calculated MLS showed reliable side identification with a mean absolute error under 1 mm, a measure error considerably far from clinical significance. Automated MLS measurement systems application can easily be foreseen as a more reliable and less interreader-susceptible tool to measure progression or response to surgical and medical treatment.

Manual measurement of hematoma volume is laborious and time-consuming in routine clinical practices. In comparison, the average analysis time of the CNN was 30 seconds in our cohort. As decisions regarding subsequent interventions are made on the basis of a few key features,18,19 AI-based systems can accelerate decision-making and can reckon useful information often overlooked in busy clinical practices or nondedicated neuroradiology departments as well as streamline patient transfers and treatments in community/referral hospitals. Moreover, the use of computer-aided diagnosis systems in academic centers can guide the initial read, usually performed by junior readers, minimizing misinterpretations and therapeutic consequences.20,21 Manual and automated segmentation agreement and bias are easily visualized with the Bland-Altman plot.

AI-guided triaging of ischemic stroke with automated detection of LVO is integrated in the workflow of many large institutions, as it facilitates the notification of critical findings to the treatment team and has been shown not only to reduce time to treatment but also to improve patient outcomes.22 Applications for the automated detection of intracranial hemorrhage are available in commercial software packages and are already in use in clinical practice (Aidoc; Rapid ICH, RapidAi, ISchemaView Inc.; Viz ICH, Viz.ai).

Apart from clinical applications, automated measurement and segmentation may prove useful in the population prevalence analysis of large data sets or for the screening of inclusion criteria in clinical trials such as those used for the EMBOLISE (Embolization of the Middle Meningeal Artery With Onyx Liquid Embolic System for Subacute and Chronic Subdural Hematoma)23 and MEMBRANE (Middle Meningeal Artery Embolization for the Treatment of Subdural Hematomas With Trufill n-BCA) trials.24 CNN use for automated trial selection and recruitment (Viz ICH) is currently in use for the ENRICH (Early Minimally Invasive Removal of Intracerebral Hemorrhage)25 trial.

The development of similar AI-based platforms for the detection and segmentation of SDH has already been proposed.2628 However, most of those methods lack proper external validation or prospective clinical applications. Yu et al.29 recently described a promising novel dimension-reduced deep learning segmentation method (DR-UNet) showing performances similar to expert clinicians in different types of hematomas, including 11 SDHs. Similar to the automated LVO stroke selection and ICH detection mentioned above, we foresee a similar potential application of automated SDH segmentation and MLS analysis for streamlining the selection and management of potential surgical candidates.30

The Viz.ai SDH CNN showed optimal sensitivity, with a PPV of 90.1%, which makes it a promising diagnostic tool. Minimizing spectrum bias with the use of an external data set, completely blinded to model development, is a known methodological resource for validation of the CNN for medical diagnostic purposes.31 Furthermore, our testing data set included different populations usually encountered in clinical practice and hematomas localized in both the supra- and infratentorial compartments. Its reliability with multiple scanners by different manufacturers is another advantage not to be overlooked.

This study has several limitations. First, the CNN was assessed on a retrospective single-center cohort. Second, 27 examinations were considered technically inadequate and thus were excluded from the analysis. Third, the assessment of MLS in the test group could be limited, as there were only 12 cases with MLS above 3 mm. Moreover, our data set did not include large SDHs with complete deformation of midline structures, which is a known pitfall in automated MLS measurements.2 Another point is that the method used for the measurement of maximum thickness relies solely on axial measurements; however, in real clinical settings, neuroradiologists and dedicated clinicians often use coronal and sagittal reconstructions.

The aim of our work is to lay the foundation for future development of the CNN including more complex assessments with additional features. The next iterations of the Viz.ai SDH CNN should focus on automated SDH type analysis and discrimination, longitudinal evaluation of progression or treatment response of patients, as well as system performance with concomitant findings such as intracranial tumors, ischemic strokes, or postoperative changes. Future directions should go far beyond the simple image analysis task, and machine learning algorithms should also focus on the correlation with patient outcomes and individualized predicted models.

Conclusions

The Viz.ai SDH CNN performed exceptionally well at identifying and quantifying key features of SDH in an independent validation imaging data set.

Acknowledgments

Dr. Kan has received grant support from the NIH (grant no. 1U1EB029353-01) and Joe Niekro Foundation (grant no. CON30914) for non–study-related clinical or research effort.

Disclosures

Dr. Kan is a consultant for Stryker Neurovascular, Imperative Care, Cerenovus, and MicroVention and has received grant support from Siemens for non–study-related clinical or research effort.

Author Contributions

Conception and design: Kan. Acquisition of data: Colasurdo, Leibushor, Robledo, Vasandani, Luna, Rao, Garcia, Srinivasan. Analysis and interpretation of data: Kan, Colasurdo, Sheth. Drafting the article: Colasurdo. Critically revising the article: Kan. Reviewed submitted version of manuscript: all authors. Statistical analysis: Colasurdo. Administrative/technical/material support: Garcia. Study supervision: Kan.

Supplemental Information

Abstract Presentations

Portions of this work were accepted as a poster presentation at the 2022 Congress of Neurological Surgeons Annual Meeting to be held in San Francisco, California, on October 8–12, 2022.

References

  • 1

    Nakaguchi H, Tanishima T, Yoshimasu N. Factors in the natural history of chronic subdural hematomas that influence their postoperative recurrence. J Neurosurg. 2001;95(2):256262.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 2

    Yang W, Huang J. Chronic subdural hematoma: epidemiology and natural history. Neurosurg Clin N Am. 2017;28(2):205210.

  • 3

    Feghali J, Yang W, Huang J. Updates in chronic subdural hematoma: epidemiology, etiology, pathogenesis, treatment, and outcome. World Neurosurg. 2020;141:339345.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 4

    Kolias AG, Chari A, Santarius T, Hutchinson PJ. Chronic subdural haematoma: modern management and emerging therapies. Nat Rev Neurol. 2014;10(10):570578.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 5

    Weigel R, Schmiedek P, Krauss JK. Outcome of contemporary surgery for chronic subdural haematoma: evidence based review. J Neurol Neurosurg Psychiatry. 2003;74(7):937943.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 6

    Ducruet AF, Grobelny BT, Zacharia BE, et al. The surgical management of chronic subdural hematoma. Neurosurg Rev. 2012;35(2):155169.

  • 7

    Lee JJ, Segar DJ, Morrison JF, Mangham WM, Lee S, Asaad WF. Subdural hematoma as a major determinant of short-term outcomes in traumatic brain injury. J Neurosurg. 2018;128(1):236249.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 8

    MacKenzie EJ. Epidemiology of injuries: current trends and future challenges. Epidemiol Rev. 2000;22(1):112119.

  • 9

    Servadei F. Prognostic factors in severely head injured adult patients with acute subdural haematoma’s. Acta Neurochir (Wien). 1997;139(4):279285.

  • 10

    Song C, Ren X, Zhao B, Fu H, Lin S, Zhang Y. Analysis of prognostic factors for patients with traumatic acute subdural hematomas treated by surgery. Article in Chinese. Zhonghua Yi Xue Za Zhi. 2014;94(17):13491352.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 11

    Atalay T, Ak H, Gülsen I, Karacabey S. Risk factors associated with mortality and survival of acute subdural hematoma: A retrospective study. J Res Med Sci. 2019;24:27.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 12

    Murray NM, Unberath M, Hager GD, Hui FK. Artificial intelligence to diagnose ischemic stroke and identify large vessel occlusions: a systematic review. J Neurointerv Surg. 2020;12(2):156164.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 13

    Soun JE, Chow DS, Nagamine M, et al. Artificial intelligence and acute stroke imaging. AJNR Am J Neuroradiol. 2021;42(1):211.

  • 14

    Ginat D. Implementation of machine learning software on the radiology worklist decreases scan view delay for the detection of intracranial hemorrhage on CT. Brain Sci. 2021;11(7):832.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 15

    Yushkevich PA, Piven J, Hazlett HC, et al. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage. 2006;31(3):11161128.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 16

    Wilson EB. Probable inference, the law of succession, and statistical inference. J Am Stat Assoc. 1927;22(158):209212.

  • 17

    Rice ME, Harris GT. Comparing effect sizes in follow-up studies: ROC area, Cohen’s d, and r. Law Hum Behav. 2005;29(5):615620.

  • 18

    Marshall LF, Marshall SB, Klauber MR, et al. The diagnosis of head injury requires a classification based on computed axial tomography. J Neurotrauma. 1992;9(suppl 1):S287S292.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 19

    Bajsarowicz P, Prakash I, Lamoureux J, et al. Nonsurgical acute traumatic subdural hematoma: what is the risk?. J Neurosurg. 2015;123(5):11761183.

  • 20

    Strub WM, Leach JL, Tomsick T, Vagal A. Overnight preliminary head CT interpretations provided by residents: locations of misidentified intracranial hemorrhage. AJNR Am J Neuroradiol. 2007;28(9):16791682.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 21

    Lal NR, Murray UM, Eldevik OP, Desmond JS. Clinical consequences of misinterpretations of neuroradiologic CT scans by on-call radiology residents. AJNR Am J Neuroradiol. 2000;21(1):124129.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 22

    Hassan AE, Ringheanu VM, Rabah RR, Preston L, Tekle WG, Qureshi AI. Early experience utilizing artificial intelligence shows significant reduction in transfer times and length of stay in a hub and spoke model. Interv Neuroradiol. 2020;26(5):615622.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 23

    Medtronic Neurovascular Clinical Affairs. A Study of the Embolization of the Middle Meningeal Artery With ONYX™ Liquid Embolic System in the Treatment of Subacute and Chronic Subdural Hematoma (EMBOLISE). Accessed August 23, 2022. https://clinicaltrials.gov/ct2/show/NCT04402632

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 24

    Cerenovus, Part of DePuy Synthes Products, Inc. Middle Meningeal Artery Embolization for the Treatment of Subdural Hematomas With TRUFILL® N-BCA. Accessed August 23, 2022. https://clinicaltrials.gov/ct2/show/NCT04816591

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 25

    Nico Corporation. ENRICH: A Multi-Center, Randomized, Clinical Trial Comparing Standard Medical Management to Early Surgical Hematoma Evacuation Using Minimally Invasive Parafascicular Surgery (MIPS) in the Treatment of Intracerebral Hemorrhage (ICH). Accessed August 23, 2022. https://clinicaltrials.gov/ct2/show/NCT02880878

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 26

    Farzaneh N, Williamson CA, Jiang C, et al. Automated segmentation and severity analysis of subdural hematoma for patients with traumatic brain injuries. Diagnostics (Basel). 2020;10(10):E773.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 27

    Kellogg RT, Vargas J, Barros G, et al. Segmentation of chronic subdural hematomas using 3D convolutional neural networks. World Neurosurg. 2021;148:e58e65.

  • 28

    Vidhya V, Gudigar A, Raghavendra U, et al. Automated detection and screening of traumatic brain injury (TBI) using computed tomography images: a comprehensive review and future perspectives. Int J Environ Res Public Health. 2021;18(12):6499.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 29

    Yu N, Yu H, Li H, Ma N, Hu C, Wang J. A robust deep learning segmentation method for hematoma volumetric detection in intracerebral hemorrhage. Stroke. 2022;53:167176.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 30

    Kan P, Maragkos GA, Srivatsan A, et al. Middle meningeal artery embolization for chronic subdural hematoma: a multi-center experience of 154 consecutive embolizations. Neurosurgery. 2021;88(2):268277.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 31

    Park SH, Han K. Methodologic guide for evaluating clinical performance and effect of artificial intelligence technology for medical diagnosis and prediction. Radiology. 2018;286(3):800809.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • Collapse
  • Expand

A mosaic of photos submitted by female neurosurgeons, overlaid with the shape of the female symbol. The mosaic is a demonstration of a crucial but underrepresented population in the neurosurgery workforce. © Sarah Woodrow, published with permission. See the article by Mulligan et al. (pp 1088–1097).

  • FIG. 1.

    Diagram illustrating the included and excluded cases in the analyzed data sets.

  • FIG. 2.

    Receiver operating characteristic curve analysis plot based on SDH volume showing an area under the curve (AUC) above 95%. Figure is available in color online only.

  • FIG. 3.

    Pearson correlation coefficient (A) and Bland-Altman plot (B) comparing manual and automated volume segmentations. Figure is available in color online only.

  • FIG. 4.

    A: Example of algorithm output in a case with bilateral SDHs, with volume and maximum thickness measurements. B: Additional example of segmentation and maximum thickness measurement in a patient with a left-sided acute SDH, a left frontal intraparenchymal hemorrhage, and ipsilateral frontoparietal subarachnoid hemorrhage. Figure is available in color online only.

  • FIG. 5.

    MLS deviation feature in a patient with bilateral SDHs (A) and right-sided SDH (B). Additional example of segmentation and maximum thickness measurement in a patient with an SDH along the falx cerebri (C). Common diagnostic pitfalls flagged as an SDH are enlarged pericephalic subarachnoid spaces (D). Hyperdense asymmetrical pericerebral subarachnoid spaces (E) and hyperdense tentorial leaflets (F) can be additional examples of false-positive flagging. Figure is available in color online only.

  • 1

    Nakaguchi H, Tanishima T, Yoshimasu N. Factors in the natural history of chronic subdural hematomas that influence their postoperative recurrence. J Neurosurg. 2001;95(2):256262.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 2

    Yang W, Huang J. Chronic subdural hematoma: epidemiology and natural history. Neurosurg Clin N Am. 2017;28(2):205210.

  • 3

    Feghali J, Yang W, Huang J. Updates in chronic subdural hematoma: epidemiology, etiology, pathogenesis, treatment, and outcome. World Neurosurg. 2020;141:339345.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 4

    Kolias AG, Chari A, Santarius T, Hutchinson PJ. Chronic subdural haematoma: modern management and emerging therapies. Nat Rev Neurol. 2014;10(10):570578.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 5

    Weigel R, Schmiedek P, Krauss JK. Outcome of contemporary surgery for chronic subdural haematoma: evidence based review. J Neurol Neurosurg Psychiatry. 2003;74(7):937943.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 6

    Ducruet AF, Grobelny BT, Zacharia BE, et al. The surgical management of chronic subdural hematoma. Neurosurg Rev. 2012;35(2):155169.

  • 7

    Lee JJ, Segar DJ, Morrison JF, Mangham WM, Lee S, Asaad WF. Subdural hematoma as a major determinant of short-term outcomes in traumatic brain injury. J Neurosurg. 2018;128(1):236249.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 8

    MacKenzie EJ. Epidemiology of injuries: current trends and future challenges. Epidemiol Rev. 2000;22(1):112119.

  • 9

    Servadei F. Prognostic factors in severely head injured adult patients with acute subdural haematoma’s. Acta Neurochir (Wien). 1997;139(4):279285.

  • 10

    Song C, Ren X, Zhao B, Fu H, Lin S, Zhang Y. Analysis of prognostic factors for patients with traumatic acute subdural hematomas treated by surgery. Article in Chinese. Zhonghua Yi Xue Za Zhi. 2014;94(17):13491352.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 11

    Atalay T, Ak H, Gülsen I, Karacabey S. Risk factors associated with mortality and survival of acute subdural hematoma: A retrospective study. J Res Med Sci. 2019;24:27.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 12

    Murray NM, Unberath M, Hager GD, Hui FK. Artificial intelligence to diagnose ischemic stroke and identify large vessel occlusions: a systematic review. J Neurointerv Surg. 2020;12(2):156164.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 13

    Soun JE, Chow DS, Nagamine M, et al. Artificial intelligence and acute stroke imaging. AJNR Am J Neuroradiol. 2021;42(1):211.

  • 14

    Ginat D. Implementation of machine learning software on the radiology worklist decreases scan view delay for the detection of intracranial hemorrhage on CT. Brain Sci. 2021;11(7):832.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 15

    Yushkevich PA, Piven J, Hazlett HC, et al. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage. 2006;31(3):11161128.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 16

    Wilson EB. Probable inference, the law of succession, and statistical inference. J Am Stat Assoc. 1927;22(158):209212.

  • 17

    Rice ME, Harris GT. Comparing effect sizes in follow-up studies: ROC area, Cohen’s d, and r. Law Hum Behav. 2005;29(5):615620.

  • 18

    Marshall LF, Marshall SB, Klauber MR, et al. The diagnosis of head injury requires a classification based on computed axial tomography. J Neurotrauma. 1992;9(suppl 1):S287S292.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 19

    Bajsarowicz P, Prakash I, Lamoureux J, et al. Nonsurgical acute traumatic subdural hematoma: what is the risk?. J Neurosurg. 2015;123(5):11761183.

  • 20

    Strub WM, Leach JL, Tomsick T, Vagal A. Overnight preliminary head CT interpretations provided by residents: locations of misidentified intracranial hemorrhage. AJNR Am J Neuroradiol. 2007;28(9):16791682.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 21

    Lal NR, Murray UM, Eldevik OP, Desmond JS. Clinical consequences of misinterpretations of neuroradiologic CT scans by on-call radiology residents. AJNR Am J Neuroradiol. 2000;21(1):124129.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 22

    Hassan AE, Ringheanu VM, Rabah RR, Preston L, Tekle WG, Qureshi AI. Early experience utilizing artificial intelligence shows significant reduction in transfer times and length of stay in a hub and spoke model. Interv Neuroradiol. 2020;26(5):615622.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 23

    Medtronic Neurovascular Clinical Affairs. A Study of the Embolization of the Middle Meningeal Artery With ONYX™ Liquid Embolic System in the Treatment of Subacute and Chronic Subdural Hematoma (EMBOLISE). Accessed August 23, 2022. https://clinicaltrials.gov/ct2/show/NCT04402632

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 24

    Cerenovus, Part of DePuy Synthes Products, Inc. Middle Meningeal Artery Embolization for the Treatment of Subdural Hematomas With TRUFILL® N-BCA. Accessed August 23, 2022. https://clinicaltrials.gov/ct2/show/NCT04816591

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 25

    Nico Corporation. ENRICH: A Multi-Center, Randomized, Clinical Trial Comparing Standard Medical Management to Early Surgical Hematoma Evacuation Using Minimally Invasive Parafascicular Surgery (MIPS) in the Treatment of Intracerebral Hemorrhage (ICH). Accessed August 23, 2022. https://clinicaltrials.gov/ct2/show/NCT02880878

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 26

    Farzaneh N, Williamson CA, Jiang C, et al. Automated segmentation and severity analysis of subdural hematoma for patients with traumatic brain injuries. Diagnostics (Basel). 2020;10(10):E773.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 27

    Kellogg RT, Vargas J, Barros G, et al. Segmentation of chronic subdural hematomas using 3D convolutional neural networks. World Neurosurg. 2021;148:e58e65.

  • 28

    Vidhya V, Gudigar A, Raghavendra U, et al. Automated detection and screening of traumatic brain injury (TBI) using computed tomography images: a comprehensive review and future perspectives. Int J Environ Res Public Health. 2021;18(12):6499.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 29

    Yu N, Yu H, Li H, Ma N, Hu C, Wang J. A robust deep learning segmentation method for hematoma volumetric detection in intracerebral hemorrhage. Stroke. 2022;53:167176.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 30

    Kan P, Maragkos GA, Srivatsan A, et al. Middle meningeal artery embolization for chronic subdural hematoma: a multi-center experience of 154 consecutive embolizations. Neurosurgery. 2021;88(2):268277.

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 31

    Park SH, Han K. Methodologic guide for evaluating clinical performance and effect of artificial intelligence technology for medical diagnosis and prediction. Radiology. 2018;286(3):800809.

    • PubMed
    • Search Google Scholar
    • Export Citation

Metrics

All Time Past Year Past 30 Days
Abstract Views 4037 112 0
Full Text Views 1451 1057 60
PDF Downloads 1256 784 71
EPUB Downloads 0 0 0