Reformatted images improve the detection rate of acute traumatic subdural hematomas on brain CT compared with axial images alone

Application of deep learning models for detection of subdural hematoma: a systematic review and meta-analysis

BACKGROUND

This study aimed to investigate the application of deep learning (DL) models for the detection of subdural hematoma (SDH).

METHODS

We conducted a comprehensive search using relevant keywords. Articles extracted were original studies in which sensitivity and/or specificity were reported. Two different approaches of frequentist and Bayesian inference were applied. For quality and risk of bias assessment we used Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2).

RESULTS

We analyzed 22 articles that included 1,997,749 patients. In the first step, the frequentist method showed a pooled sensitivity of 88.8% (95% confidence interval (CI): 83.9% to 92.4%) and a specificity of 97.2% (95% CI 94.6% to 98.6%). In the second step, using Bayesian methods including 11 studies that reported sensitivity and specificity, a sensitivity rate of 86.8% (95% CI: 77.6% to 92.9%) at a specificity level of 86.9% (95% CI: 60.9% to 97.2%) was achieved. The risk of bias assessment was not remarkable using QUADAS-2.

CONCLUSION

DL models might be an appropriate tool for detecting SDHs with a reasonably high sensitivity and specificity.

A Quantitative Subarachnoid Hemorrhage Grading System, Including Supratentorial and Infratentorial Cisterns, With Multiplanar Computed Tomography Reformations

Background

Subarachnoid hemorrhage (SAH) grading scales typically evaluate a limited number of cisterns on the axial plane. The goal of our study is to apply a simple quantitative yet comprehensive SAH grading scale to all major intracranial cisterns, including the infratentorial cisterns, with multiplanar computed tomography (CT) reformations.

Methodology

We performed a retrospective review of 94 consecutive cases of spontaneous SAH presenting within 72 hours of onset. SAH was categorized into five grades based on the short-axis thickness of SAH in 20 intracranial cisterns measured on the axial, coronal, and sagittal planes. Statistical analysis was performed for inter-rater agreement with kappa statistics, for inter-plane agreement by Spearman correlation statistics, and for inter-rater and inter-plane agreement by Pearson correlation statistics.

Results

The extended kappa coefficient for the three reviewers across all 20 cisterns varied from 0.38 (0.27, 0.50) to 0.59 (0.52, 0.65) on the axial plane. The kappa coefficient for two reviewers varied from 0.46 (0.33, 0.59) to 0.70 (0.60, 0.80) on the coronal plane and from 0.35 (0.20, 0.49) to 0.87 (0.77, 0.96) on the sagittal plane. The average grade of cisterns per case demonstrated mostly excellent correlation between the imaging planes with Spearman correlation statistics (≥0.70). Pairwise concordance correlation coefficient of the total SAH score revealed agreement ranging from 0.81 to 0.90 in all three planes. Pearson correlation statistics of the average total SAH scores revealed excellent correlation among the three planes (≥0.91).

Conclusion

A simple quantitative SAH grading scale can be successfully applied to the supratentorial and infratentorial cisterns in three standard CT imaging planes.

Multiplanar CT evaluation of aneurysm rupture signs in subarachnoid hemorrhage

PURPOSE

In subarachnoid hemorrhage, noncontrast CT features are used to guide the localization of ruptured aneurysms on CT angiography and DSA. Multiplanar CT may improve the localization of aneurysm rupture sites over axial plane CT alone.

METHODS

Multiplanar CT in three orthogonal planes was used to evaluate 94 cases of SAH. Two investigators independently evaluated each imaging plane for focal thick SAH with mass effect, intracerebral hemorrhage, focal edema, filling defect, subdural hemorrhage, and dominant intraventricular hemorrhage. Also, rupture site was qualitatively identified by combining these variables in each plane and combination of three planes. DSA served as the gold standard to locate the rupture sites.

RESULTS

Interobserver agreement was k 0.60 to 0.79 for axial, k 0.43 to 0.86 for coronal and k 0.43 to 0.74 for sagittal planes. Good to substantial agreement was observed for the localization of rupture site in three planes (focal SAH with mass effect - k 0.78 to 0.85; filling defect - k 0.95 to 1.0; intracerebral hemorrhage - k 1.0; focal edema k 1.0; subdural hemorrhage - k 0.61 to 0.83). Dominant intraventricular hemorrhage revealed significant association with DSA to locate ruptured aneurysms (Fisher's exact test - Pr <  = P (< 0.001)). With non-missing data, frequency of correct ratings to locate rupture site was 66/67 (99%) in axial plane, 59/66 (89%) in coronal plane, 64/67 (96%) in sagittal plane and 77/77 (100%) in combined 3 planes.

CONCLUSIONS

Multiplanar CT head is more successful than axial plane CT alone for the localization of aneurysm rupture sites in SAH.

Deep learning algorithm in detecting intracranial hemorrhages on emergency computed tomographies

BACKGROUND

Highly accurate detection of intracranial hemorrhages (ICH) on head computed tomography (HCT) scans can prove challenging at high-volume centers. This study aimed to determine the number of additional ICHs detected by an artificial intelligence (AI) algorithm and to evaluate reasons for erroneous results at a level I trauma center with teleradiology services.

METHODS

In a retrospective multi-center cohort study, consecutive emergency non-contrast HCT scans were analyzed by a commercially available ICH detection software (AIDOC, Tel Aviv, Israel). Discrepancies between AI analysis and initial radiology report (RR) were reviewed by a blinded neuroradiologist to determine the number of additional ICHs detected and evaluate reasons leading to errors.

RESULTS

4946 HCT (05/2020-09/2020) from 18 hospitals were included in the analysis. 205 reports (4.1%) were classified as hemorrhages by both radiology report and AI. Out of a total of 162 (3.3%) discrepant reports, 62 were confirmed as hemorrhages by the reference neuroradiologist. 33 ICHs were identified exclusively via RRs. The AI algorithm detected an additional 29 instances of ICH, missed 12.4% of ICH and overcalled 1.9%; RRs missed 10.9% of ICHs and overcalled 0.2%. Many of the ICHs missed by the AI algorithm were located in the subarachnoid space (42.4%) and under the calvaria (48.5%). 85% of ICHs missed by RRs occurred outside of regular working-hours. Calcifications (39.3%), beam-hardening artifacts (18%), tumors (15.7%), and blood vessels (7.9%) were the most common reasons for AI overcalls. ICH size, image quality, and primary examiner experience were not found to be significantly associated with likelihood of incorrect AI results.

CONCLUSION

Complementing human expertise with AI resulted in a 12.2% increase in ICH detection. The AI algorithm overcalled 1.9% HCT.

TRIAL REGISTRATION

German Clinical Trials Register (DRKS-ID: DRKS00023593).

Automated detection and segmentation of intracranial hemorrhage suspect hyperdensities in non-contrast-enhanced CT scans of acute stroke patients

Objectives

Artif icial intelligence (AI)–based image analysis is increasingly applied in the acute stroke field. Its implementation for the detection and quantification of hemorrhage suspect hyperdensities in non-contrast-enhanced head CT (NCCT) scans may facilitate clinical decision-making and accelerate stroke management.

Methods

NCCTs of 160 patients with suspected acute stroke were analyzed regarding the presence or absence of acute intracranial hemorrhages (ICH) using a novel AI-based algorithm. Read was performed by two blinded neuroradiology residents (R1 and R2). Ground truth was established by an expert neuroradiologist. Specificity, sensitivity, and area under the curve were calculated for ICH and intraparenchymal hemorrhage (IPH) detection. IPH-volumes were segmented and quantified automatically by the algorithm and semi-automatically. Intraclass correlation coefficient (ICC) and Dice coefficient (DC) were calculated.

Results

In total, 79 of 160 patients showed acute ICH, while 47 had IPH. Sensitivity and specificity for ICH detection were 0.91 and 0.89 for the algorithm; 0.99 and 0.98 for R1; and 1.00 and 0.98 for R2. Sensitivity and specificity for IPH detection were 0.98 and 0.89 for the algorithm; 0.83 and 0.99 for R1; and 0.91 and 0.99 for R2. Interreader reliability for ICH and IPH detection showed strong agreements for the algorithm (0.80 and 0.84), R1 (0.96 and 0.84), and R2 (0.98 and 0.92), respectively. ICC indicated an excellent (0.98) agreement between the algorithm and the reference standard of the IPH-volumes. The mean DC was 0.82.

Conclusion

The AI-based algorithm reliably assessed the presence or absence of acute ICHs in this dataset and quantified IPH volumes precisely.

Key Points

• Artificial intelligence (AI) is able to detect hyperdense volumes on brain CTs reliably.

• Sensitivity and specificity are highest for the detection of intraparenchymal hemorrhages.

• Interreader reliability for hemorrhage detection shows strong agreement for AI and human readers.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00330-021-08352-4.

An experimental study on the early diagnosis of traumatic brain injury in rabbits based on a noncontact and portable system

Closed cerebral hemorrhage (CCH) is a common symptom in traumatic brain injury (TBI) patients who suffer intracranial hemorrhage with the dura mater remaining intact. The diagnosis of CCH patients prior to hospitalization and in the early stage of the disease can help patients get earlier treatments that improve outcomes. In this study, a noncontact, portable system for early TBI-induced CCH detection was constructed that measures the magnetic induction phase shift (MIPS), which is associated with the mean brain conductivity caused by the ratio between the liquid (blood/CSF and the intracranial tissues) change. To evaluate the performance of this system, a rabbit CCH model with two severity levels was established based on the horizontal biological impactor BIM-II, whose feasibility was verified by computed tomography images of three sections and three serial slices. There were two groups involved in the experiments (group 1 with 10 TBI rabbits were simulated by hammer hit with air pressure of 600 kPa by BIM-II and group 2 with 10 TBI rabbits were simulated with 650 kPa). The MIPS values of the two groups were obtained within 30 min before and after injury. In group 1, the MIPS values showed a constant downward trend with a minimum value of −11.17 ± 2.91° at the 30th min after 600 kPa impact by BIM-II. After the 650 kPa impact, the MIPS values in group 2 showed a constant downward trend until the 25th min, with a minimum value of −16.81 ± 2.10°. Unlike group 1, the MIPS values showed an upward trend after that point. Before the injury, the MIPS values in both group 1 and group 2 did not obviously change within the 30 min measurement. Using a support vector machine at the same time point after injury, the classification accuracy of the two types of severity was shown to be beyond 90%. Combined with CCH pathological mechanisms, this system can not only achieve the detection of early functional changes in CCH but can also distinguish different severities of CCH.