Computed tomography (CT) of the head is used worldwide to diagnose neurologic emergencies. However, expertise is required to interpret these scans, and even highly trained experts may miss subtle ...life-threatening findings. For head CT, a unique challenge is to identify, with perfect or near-perfect sensitivity and very high specificity, often small subtle abnormalities on a multislice cross-sectional (three-dimensional 3D) imaging modality that is characterized by poor soft tissue contrast, low signal-to-noise using current low radiation-dose protocols, and a high incidence of artifacts. We trained a fully convolutional neural network with 4,396 head CT scans performed at the University of California at San Francisco and affiliated hospitals and compared the algorithm’s performance to that of 4 American Board of Radiology (ABR) certified radiologists on an independent test set of 200 randomly selected head CT scans. Our algorithm demonstrated the highest accuracy to date for this clinical application, with a receiver operating characteristic (ROC) area under the curve (AUC) of 0.991 ± 0.006 for identification of examinations positive for acute intracranial hemorrhage, and also exceeded the performance of 2 of 4 radiologists. We demonstrate an end-to-end network that performs joint classification and segmentation with examination-level classification comparable to experts, in addition to robust localization of abnormalities, including some that are missed by radiologists, both of which are critically important elements for this application.
In recent years, there have been major advances in deep learning algorithms for image recognition in traumatic brain injury (TBI). Interest in this area has increased due to the potential for greater ...objectivity, reduced interpretation times and, ultimately, higher accuracy. Triage algorithms that can re-order radiological reading queues have been developed, using classification to prioritize exams with suspected critical findings. Localization models move a step further to capture more granular information such as the location and, in some cases, size and subtype, of intracranial hematomas that could aid in neurosurgical management decisions. In addition to the potential to improve the clinical management of TBI patients, the use of algorithms for the interpretation of medical images may play a transformative role in enabling the integration of medical images into precision medicine. Acute TBI is one practical example that can illustrate the application of deep learning to medical imaging. This review provides an overview of computational approaches that have been proposed for the detection and characterization of acute TBI imaging abnormalities, including intracranial hemorrhage, skull fractures, intracranial mass effect, and stroke.
Objective
To determine the clinical relevance, if any, of traumatic intracranial findings on early head computed tomography (CT) and brain magnetic resonance imaging (MRI) to 3‐month outcome in mild ...traumatic brain injury (MTBI).
Methods
One hundred thirty‐five MTBI patients evaluated for acute head injury in emergency departments of 3 LEVEL I trauma centers were enrolled prospectively. In addition to admission head CT, early brain MRI was performed 12 ± 3.9 days after injury. Univariate and multivariate logistic regression were used to assess for demographic, clinical, socioeconomic, CT, and MRI features that were predictive of Extended Glasgow Outcome Scale (GOS‐E) at 3 months postinjury.
Results
Twenty‐seven percent of MTBI patients with normal admission head CT had abnormal early brain MRI. CT evidence of subarachnoid hemorrhage was associated with a multivariate odds ratio of 3.5 (p = 0.01) for poorer 3‐month outcome, after adjusting for demographic, clinical, and socioeconomic factors. One or more brain contusions on MRI, and ≥4 foci of hemorrhagic axonal injury on MRI, were each independently associated with poorer 3‐month outcome, with multivariate odds ratios of 4.5 (p = 0.01) and 3.2 (p = 0.03), respectively, after adjusting for head CT findings and demographic, clinical, and socioeconomic factors.
Interpretation
In this prospective multicenter observational study, the clinical relevance of abnormal findings on early brain imaging after MTBI is demonstrated. The addition of early CT and MRI markers to a prognostic model based on previously known demographic, clinical, and socioeconomic predictors resulted in a >2‐fold increase in the explained variance in 3‐month GOS‐E. ANN NEUROL 2013;73:224–235
Annually in the United States, at least 3.5 million people seek medical attention for traumatic brain injury (TBI). The development of therapies for TBI is limited by the absence of diagnostic and ...prognostic biomarkers. Microtubule-associated protein tau is an axonal phosphoprotein. To date, the presence of the hypophosphorylated tau protein (P-tau) in plasma from patients with acute TBI and chronic TBI has not been investigated.
To examine the associations between plasma P-tau and total-tau (T-tau) levels and injury presence, severity, type of pathoanatomic lesion (neuroimaging), and patient outcomes in acute and chronic TBI.
In the TRACK-TBI Pilot study, plasma was collected at a single time point from 196 patients with acute TBI admitted to 3 level I trauma centers (<24 hours after injury) and 21 patients with TBI admitted to inpatient rehabilitation units (mean SD, 176.4 44.5 days after injury). Control samples were purchased from a commercial vendor. The TRACK-TBI Pilot study was conducted from April 1, 2010, to June 30, 2012. Data analysis for the current investigation was performed from August 1, 2015, to March 13, 2017.
Plasma samples were assayed for P-tau (using an antibody that specifically recognizes phosphothreonine-231) and T-tau using ultra-high sensitivity laser-based immunoassay multi-arrayed fiberoptics conjugated with rolling circle amplification.
In the 217 patients with TBI, 161 (74.2%) were men; mean (SD) age was 42.5 (18.1) years. The P-tau and T-tau levels and P-tau-T-tau ratio in patients with acute TBI were higher than those in healthy controls. Receiver operating characteristic analysis for the 3 tau indices demonstrated accuracy with area under the curve (AUC) of 1.000, 0.916, and 1.000, respectively, for discriminating mild TBI (Glasgow Coma Scale GCS score, 13-15, n = 162) from healthy controls. The P-tau level and P-tau-T-tau ratio were higher in individuals with more severe TBI (GCS, ≤12 vs 13-15). The P-tau level and P-tau-T-tau ratio outperformed the T-tau level in distinguishing cranial computed tomography-positive from -negative cases (AUC = 0.921, 0.923, and 0.646, respectively). Acute P-tau levels and P-tau-T-tau ratio weakly distinguished patients with TBI who had good outcomes (Glasgow Outcome Scale-Extended GOS-E, 7-8) (AUC = 0.663 and 0.658, respectively) and identified those with poor outcomes (GOS-E, ≤4 vs >4) (AUC = 0.771 and 0.777, respectively). Plasma samples from patients with chronic TBI also showed elevated P-tau levels and a P-tau-T-tau ratio significantly higher than that of healthy controls, with both P-tau indices strongly discriminating patients with chronic TBI from healthy controls (AUC = 1.000 and 0.963, respectively).
Plasma P-tau levels and P-tau-T-tau ratio outperformed T-tau level as diagnostic and prognostic biomarkers for acute TBI. Compared with T-tau levels alone, P-tau levels and P-tau-T-tau ratios show more robust and sustained elevations among patients with chronic TBI.
Biomarkers are important for accurate diagnosis of complex disorders such as traumatic brain injury (TBI). For a complex and multifaceted condition such as TBI, it is likely that a single biomarker ...will not reflect the full spectrum of the response of brain tissue to injury. Ubiquitin C-terminal hydrolase L1 (UCH-L1) and glial fibrillary acidic protein (GFAP) are among of the most widely studied biomarkers for TBI. Because UCH-L1 and GFAP measure distinct molecular events, we hypothesized that analysis of both biomarkers would be superior to analysis of each alone for the diagnosis and prognosis of TBI. Serum levels of UCH-L1 and GFAP were measured in a cohort of 206 patients with TBI enrolled in a multicenter observational study (Transforming Research and Clinical Knowledge in Traumatic Brain Injury TRACK-TBI). Levels of the two biomarkers were weakly correlated to each other (r=0.364). Each biomarker in isolation had good sensitivity and sensitivity for discriminating between TBI patients and healthy controls (area under the curve AUC 0.87 and 0.91 for UCH-L1 and GFAP, respectively). When biomarkers were combined, superior sensitivity and specificity for diagnosing TBI was obtained (AUC 0.94). Both biomarkers discriminated between TBI patients with intracranial lesions on CT scan and those without such lesions, but GFAP measures were significantly more sensitive and specific (AUC 0.88 vs. 0.71 for UCH-L1). For association with outcome 3 months after injury, neither biomarker had adequate sensitivity and specificity (AUC 0.65-0.74, for GFAP, and 0.59-0.80 for UCH-L1, depending upon Glasgow Outcome Scale Extended GOS-E threshold used). Our results support a role for multiple biomarker measurements in TBI research. ( ClinicalTrials.gov Identifier NCT01565551).
We evaluated 3T diffusion tensor imaging (DTI) for white matter injury in 76 adult mild traumatic brain injury (mTBI) patients at the semiacute stage (11.2±3.3 days), employing both whole-brain ...voxel-wise and region-of-interest (ROI) approaches. The subgroup of 32 patients with any traumatic intracranial lesion on either day-of-injury computed tomography (CT) or semiacute magnetic resonance imaging (MRI) demonstrated reduced fractional anisotropy (FA) in numerous white matter tracts, compared to 50 control subjects. In contrast, 44 CT/MRI-negative mTBI patients demonstrated no significant difference in any DTI parameter, compared to controls. To determine the clinical relevance of DTI, we evaluated correlations between 3- and 6-month outcome and imaging, demographic/socioeconomic, and clinical predictors. Statistically significant univariable predictors of 3-month Glasgow Outcome Scale-Extended (GOS-E) included MRI evidence for contusion (odds ratio OR 4.9 per unit decrease in GOS-E; p=0.01), ≥1 ROI with severely reduced FA (OR, 3.9; p=0.005), neuropsychiatric history (OR, 3.3; p=0.02), age (OR, 1.07/year; p=0.002), and years of education (OR, 0.79/year; p=0.01). Significant predictors of 6-month GOS-E included ≥1 ROI with severely reduced FA (OR, 2.7; p=0.048), neuropsychiatric history (OR, 3.7; p=0.01), and years of education (OR, 0.82/year; p=0.03). For the subset of 37 patients lacking neuropsychiatric and substance abuse history, MRI surpassed all other predictors for both 3- and 6-month outcome prediction. This is the first study to compare DTI in individual mTBI patients to conventional imaging, clinical, and demographic/socioeconomic characteristics for outcome prediction. DTI demonstrated utility in an inclusive group of patients with heterogeneous backgrounds, as well as in a subset of patients without neuropsychiatric or substance abuse history.
Brain-derived neurotrophic factor (BDNF) is important for neuronal survival and regeneration. We investigated the diagnostic and prognostic values of serum BDNF in traumatic brain injury (TBI). We ...examined serum BDNF in two independent cohorts of TBI cases presenting to the emergency departments (EDs) of the Johns Hopkins Hospital (JHH; n = 76) and San Francisco General Hospital (SFGH, n = 80), and a control group of JHH ED patients without TBI (n = 150). Findings were subsequently validated in the prospective, multi-center Transforming Research and Clinical Knowledge in TBI (TRACK-TBI) Pilot study (n = 159). We investigated the association between BDNF, glial fibrillary acidic protein (GFAP), and ubiquitin C-terminal hydrolase-L1 (UCH-L1) and recovery from TBI at 6 months in the TRACK-TBI Pilot cohort. Incomplete recovery was defined as having either post-concussive syndrome or a Glasgow Outcome Scale Extended score <8 at 6 months. Median day-of-injury BDNF concentrations (ng/mL) were lower among TBI cases (JHH TBI, 17.5 and SFGH TBI, 13.8) than in JHH controls (60.3; p = 0.0001). Among TRACK-TBI Pilot subjects, median BDNF concentrations (ng/mL) were higher in mild (8.3) than in moderate (4.3) or severe TBI (4.0; p = 0.004. In the TRACK-TBI cohort, the 75 (71.4%) subjects with very low BDNF values (i.e., <the 1st percentile for non-TBI controls, <14.2 ng/mL) had higher odds of incomplete recovery than those who did not have very low values (odds ratio, 4.0; 95% confidence interval CI: 1.5-11.0). The area under the receiver operator curve for discriminating complete and incomplete recovery was 0.65 (95% CI: 0.52-0.78) for BDNF, 0.61 (95% CI: 0.49-0.73) for GFAP, and 0.55 (95% CI: 0.43-0.66) for UCH-L1. The addition of GFAP/UCH-L1 to BDNF did not improve outcome prediction significantly. Day-of-injury serum BDNF is associated with TBI diagnosis and also provides 6-month prognostic information regarding recovery from TBI. Thus, day-of-injury BDNF values may aid in TBI risk stratification.
Plasma tau and glial fibrillary acidic protein (GFAP) are promising biomarkers for identifying traumatic brain injury (TBI) patients with intracranial trauma on computed tomography (CT). Accuracy in ...older adults with mild TBI (mTBI), the fastest growing TBI population, is unknown. Our aim was to assess for age-related differences in diagnostic accuracy of plasma tau and GFAP for identifying intracranial trauma on CT. Samples from 169 patients (age <40 years n = 79, age 40-59 years n = 60, age 60 years+ n = 30), a subset of patients from the Transforming Research and Clinical Knowledge in TBI (TRACK-TBI) Pilot study who presented with mTBI (Glasgow Coma Scale score of 13-15), received head CT, and consented to blood draw within 24 h of injury, were assayed for hyperphosphorylated-tau (P-tau), total-tau (T-tau; both via amplification-linked enhanced immunoassay using multi-arrayed fiberoptics), and GFAP (via sandwich enzyme-linked immunosorbent assay). P-tau, T-tau, P-tau:T-tau ratio, and GFAP concentration were significantly associated with CT findings. Overall, discriminative ability declined with increasing age for all assays, but this decline was only statistically significant for GFAP (area under the receiver operating characteristic curve AUC: old 0.73 reference group; ref vs. young 0.93 p = 0.037 or middle-aged 0.92 p = 0.0497). P-tau concentration consistently showed the highest diagnostic accuracy across all age-groups (AUC: old 0.84 ref vs. young 0.95 p = 0.274 or middle-aged 0.93 p = 0.367). Comparison of models including P-tau alone versus P-tau plus GFAP revealed significant added value of GFAP. In conclusion, the GFAP assay was less accurate for identifying intracranial trauma on CT among older versus younger mTBI patients. Mechanisms of this age-related difference, including role of assay methodology, specific TBI neuroanatomy, pre-existing conditions, and anti-thrombotic use, warrant further study.
Purpose To develop and evaluate a semi-supervised learning model for intracranial hemorrhage detection and segmentation on an out-of-distribution head CT evaluation set. Materials and Methods This ...retrospective study used semi-supervised learning to bootstrap performance. An initial "teacher" deep learning model was trained on 457 pixel-labeled head CT scans collected from one U.S. institution from 2010 to 2017 and used to generate pseudo labels on a separate unlabeled corpus of 25 000 examinations from the Radiological Society of North America and American Society of Neuroradiology. A second "student" model was trained on this combined pixel- and pseudo-labeled dataset. Hyperparameter tuning was performed on a validation set of 93 scans. Testing for both classification (
= 481 examinations) and segmentation (
= 23 examinations, or 529 images) was performed on CQ500, a dataset of 481 scans performed in India, to evaluate out-of-distribution generalizability. The semi-supervised model was compared with a baseline model trained on only labeled data using area under the receiver operating characteristic curve, Dice similarity coefficient, and average precision metrics. Results The semi-supervised model achieved a statistically significant higher examination area under the receiver operating characteristic curve on CQ500 compared with the baseline (0.939 95% CI: 0.938, 0.940 vs 0.907 95% CI: 0.906, 0.908;
= .009). It also achieved a higher Dice similarity coefficient (0.829 95% CI: 0.825, 0.833 vs 0.809 95% CI: 0.803, 0.812;
= .012) and pixel average precision (0.848 95% CI: 0.843, 0.853) vs 0.828 95% CI: 0.817, 0.828) compared with the baseline. Conclusion The addition of unlabeled data in a semi-supervised learning framework demonstrates stronger generalizability potential for intracranial hemorrhage detection and segmentation compared with a supervised baseline.
Semi-supervised Learning, Traumatic Brain Injury, CT, Machine Learning
Published under a CC BY 4.0 license. See also the commentary by Swimburne in this issue.
Reliable diagnosis of traumatic brain injury (TBI) is a major public health need. Glial fibrillary acidic protein (GFAP) is expressed in the central nervous system, and breakdown products (GFAP-BDP) ...are released following parenchymal brain injury. Here, we evaluate the diagnostic accuracy of elevated levels of plasma GFAP-BDP in TBI. Participants were identified as part of the prospective Transforming Research And Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) Study. Acute plasma samples (<24 h post-injury) were collected from patients presenting with brain injury who had CT imaging. The ability of GFAP-BDP level to discriminate patients with demonstrable traumatic lesions on CT, and with failure to return to pre-injury baseline at 6 months, was evaluated by the area under the receiver operating characteristic curve (AUC). Of the 215 patients included for analysis, 83% had mild, 4% had moderate, and 13% had severe TBI; 54% had acute traumatic lesions on CT. The ability of GFAP-BDP level to discriminate patients with traumatic lesions on CT as evaluated by AUC was 0.88 (95% confidence interval CI, 0.84-0.93). The optimal cutoff of 0.68 ng/mL for plasma GFAP-BDP level was associated with a 21.61 odds ratio for traumatic findings on head CT. Discriminatory ability of unfavorable 6 month outcome was lower, AUC 0.65 (95% CI, 0.55-0.74), with a 2.07 odds ratio. GFAP-BDP levels reliably distinguish the presence and severity of CT scan findings in TBI patients. Although these findings confirm and extend prior studies, a larger prospective trial is still needed to validate the use of GFAP-BDP as a routine diagnostic biomarker for patient care and clinical research. The term "mild" continues to be a misnomer for this patient population, and underscores the need for evolving classification strategies for TBI targeted therapy. (ClinicalTrials.gov number NCT01565551; NIH Grant 1RC2 NS069409).