Detection of Intracerebral Hemorrhage Using Low-Field, Portable Magnetic Resonance Imaging in Patients With Stroke.

BACKGROUND: Neuroimaging is essential for detecting spontaneous, nontraumatic intracerebral hemorrhage (ICH). Recent data suggest ICH can be characterized using low-field magnetic resonance imaging (MRI). Our primary objective was to investigate the sensitivity and specificity of ICH on a 0.064T portable MRI (pMRI) scanner using a methodology that provided clinical information to inform rater interpretations. As a secondary aim, we investigated whether the incorporation of a deep learning (DL) reconstruction algorithm affected ICH detection. METHODS: The pMRI device was deployed at Yale New Haven Hospital to examine patients presenting with stroke symptoms from October 26, 2020 to February 21, 2022. Three raters independently evaluated pMRI examinations. Raters were provided the images alongside the patient's clinical information to simulate real-world context of use. Ground truth was the closest conventional computed tomography or 1.5/3T MRI. Sensitivity and specificity results were grouped by DL and non-DL software to investigate the effects of software advances. RESULTS: A total of 189 exams (38 ICH, 89 acute ischemic stroke, 8 subarachnoid hemorrhage, 3 primary intraventricular hemorrhage, 51 no intracranial abnormality) were evaluated. Exams were correctly classified as positive or negative for ICH in 185 of 189 cases (97.9% overall accuracy). ICH was correctly detected in 35 of 38 cases (92.1% sensitivity). Ischemic stroke and no intracranial abnormality cases were correctly identified as blood-negative in 139 of 140 cases (99.3% specificity). Non-DL scans had a sensitivity and specificity for ICH of 77.8% and 97.1%, respectively. DL scans had a sensitivity and specificity for ICH of 96.6% and 99.3%, respectively. CONCLUSIONS: These results demonstrate improvements in ICH detection accuracy on pMRI that may be attributed to the integration of clinical information in rater review and the incorporation of a DL-based algorithm. The use of pMRI holds promise in providing diagnostic neuroimaging for patients with ICH.

Deep Learning Applications for Acute Stroke Management.

Brain imaging is essential to the clinical care of patients with stroke, a leading cause of disability and death worldwide. Whereas advanced neuroimaging techniques offer opportunities for aiding acute stroke management, several factors, including time delays, inter-clinician variability, and lack of systemic conglomeration of clinical information, hinder their maximal utility. Recent advances in deep machine learning (DL) offer new strategies for harnessing computational medical image analysis to inform decision making in acute stroke. We examine the current state of the field for DL models in stroke triage. First, we provide a brief, clinical practice-focused primer on DL. Next, we examine real-world examples of DL applications in pixel-wise labeling, volumetric lesion segmentation, stroke detection, and prediction of tissue fate postintervention. We evaluate recent deployments of deep neural networks and their ability to automatically select relevant clinical features for acute decision making, reduce inter-rater variability, and boost reliability in rapid neuroimaging assessments, and integrate neuroimaging with electronic medical record (EMR) data in order to support clinicians in routine and triage stroke management. Ultimately, we aim to provide a framework for critically evaluating existing automated approaches, thus equipping clinicians with the ability to understand and potentially apply DL approaches in order to address challenges in clinical practice. ANN NEUROL 2022;92:574-587.

Deep learning survival model predicts outcome after intracerebral hemorrhage from initial CT scan.

BACKGROUND: Predicting functional impairment after intracerebral hemorrhage (ICH) provides valuable information for planning of patient care and rehabilitation strategies. Current prognostic tools are limited in making long term predictions and require multiple expert-defined inputs and interpretation that make their clinical implementation challenging. This study aimed to predict long term functional impairment of ICH patients from admission non-contrast CT scans, leveraging deep learning models in a survival analysis framework. METHODS: We used the admission non-contrast CT scans from 882 patients from the Massachusetts General Hospital ICH Study for training, hyperparameter optimization, and model selection, and 146 patients from the Yale New Haven ICH Study for external validation of a deep learning model predicting functional outcome. Disability (modified Rankin scale [mRS] > 2), severe disability (mRS > 4), and dependent living status were assessed via telephone interviews after 6, 12, and 24 months. The prediction methods were evaluated by the c-index and compared with ICH score and FUNC score. RESULTS: Using non-contrast CT, our deep learning model achieved higher prediction accuracy of post-ICH dependent living, disability, and severe disability by 6, 12, and 24 months (c-index 0.742 [95% CI -0.700 to 0.778], 0.712 [95% CI -0.674 to 0.752], 0.779 [95% CI -0.733 to 0.832] respectively) compared with the ICH score (c-index 0.673 [95% CI -0.662 to 0.688], 0.647 [95% CI -0.637 to 0.661] and 0.697 [95% CI -0.675 to 0.717]) and FUNC score (c-index 0.701 [95% CI- 0.698 to 0.723], 0.668 [95% CI -0.657 to 0.680] and 0.727 [95% CI -0.708 to 0.753]). In the external independent Yale-ICH cohort, similar performance metrics were obtained for disability and severe disability (c-index 0.725 [95% CI -0.673 to 0.781] and 0.747 [95% CI -0.676 to 0.807], respectively). Similar AUC of predicting each outcome at 6 months, 1 and 2 years after ICH was achieved compared with ICH score and FUNC score. CONCLUSION: We developed a generalizable deep learning model to predict onset of dependent living and disability after ICH, which could help to guide treatment decisions, advise relatives in the acute setting, optimize rehabilitation strategies, and anticipate long-term care needs.

Poor Oral Health Is Associated with Worse Brain Imaging Profiles.

Importance: Poor oral health is a modifiable risk factor that is associated with a variety of health outcomes. However, the relationship between oral and brain health is not well understood. Objective: To test the hypothesis that poor oral health is associated with worse neuroimaging brain health profiles in persons without stroke or dementia. Design: We conducted a 2-stage cross-sectional neuroimaging study using data from the UK Biobank (UKB). First, we tested for association between self-reported poor oral health and MRI neuroimaging markers of brain health. Second, we used Mendelian Randomization (MR) analyses to test for association between genetically-determined poor oral health and the same neuroimaging markers. Setting: Ongoing population study in the United Kingdom. The UKB enrolled participants between 2006 and 2010. Data analysis was performed from September 1, 2022, to January 10, 2023. Participants: 40,175 persons aged 40 to 70 enrolled between 2006 to 2010 who underwent a dedicated research brain MRI between 2012 and 2013. Exposures: During MRI assessment, poor oral health was defined as the presence of dentures or loose teeth. As instruments for the MR analysis, we used 116 independent DNA sequence variants known to significantly increase the composite risk of decayed, missing, or filled teeth and dentures. Main Outcomes and Measures: As neuroimaging markers of brain health, we assessed the volume of white matter hyperintensities (WMH), as well as aggregate measures of fractional anisotropy (FA) and mean diffusivity (MD), two metrics indicative of white matter tract disintegrity obtained through diffusion tensor imaging. These measurements were evaluated across 48 distinct brain regions, with FA and MD values for each region also considered as individual outcomes for the MR method. Results: Among study participants, 5,470 (14%) had poor oral health. We found that poor oral health was associated with a 9% increase in WMH volume (beta = 0.09, standard deviation (SD) = 0.014, p P< 0.001), a 10% change in the aggregate FA score (beta = 0.10, SD = 0.013, P < 0.001), and a 5% change in the aggregate MD score (beta = 0.05, SD = 0.013, P < 0.001). Genetically-determined poor oral health was associated with a 30% increase in WMH volume (beta = 0.30, SD = 0.06, P < 0.001), a 43% change in aggregate FA score (beta = 0.42, SD = 0.06, P < 0.001), and an 10% change in aggregate MD score (beta = 0.10, SD = 0.03, P = 0.01). Conclusions and Relevance: Among middle age Britons without stroke or dementia enrolled in a large population study, poor oral health was associated with worse neuroimaging brain health profiles. Genetic analyses confirmed these associations, supporting a potential causal association. Because the neuroimaging markers evaluated in the current study are established risk factors for stroke and dementia, our results suggest that oral health may be a promising target for interventions focused on improving brain health.

Polygenic Susceptibility to Hypertension and Cognitive Performance in Middle-aged Persons Without Stroke or Dementia.

BACKGROUND AND OBJECTIVES: Mounting evidence indicates that hypertension leads to a higher risk of dementia. Hypertension is a highly heritable trait, and a higher polygenic susceptibility to hypertension (PSH) is known to associate with a higher risk of dementia. We tested the hypothesis that a higher PSH leads to worse cognitive performance in middle-aged persons without dementia. Confirming this hypothesis would support follow-up research focused on using hypertension-related genomic information to risk-stratify middle-aged adults before hypertension develops. METHODS: We conducted a nested cross-sectional genetic study within the UK Biobank (UKB). Study participants with a history of dementia or stroke were excluded. We categorized participants as having low (≤20th percentile), intermediate, or high (≥80th percentile) PSH according to results of 2 polygenic risk scores for systolic and diastolic blood pressure (BP) generated with data on 732 genetic risk variants. A general cognitive ability score was calculated as the first component of an analysis that included the results of 5 cognitive tests. Primary analyses focused on Europeans, and secondary analyses included all race/ethnic groups. RESULTS: Of the 502,422 participants enrolled in the UKB, 48,118 (9.6%) completed the cognitive evaluation, including 42,011 (8.4%) of European ancestry. Multivariable regression models using systolic BP-related genetic variants indicated that compared with study participants with a low PSH, those with intermediate and high PSH had reductions of 3.9% (ß -0.039, SE 0.012) and 6.6% (ß -0.066, SE 0.014), respectively, in their general cognitive ability score (p < 0.001). Secondary analyses including all race/ethnic groups and using diastolic BP-related genetic variants yielded similar results (p < 0.05 for all tests). Analyses evaluating each cognitive test separately indicated that reaction time, numeric memory, and fluid intelligence drove the association between PSH and general cognitive ability score (all individual tests, p < 0.05). DISCUSSION: Among nondemented, community-dwelling, middle-aged Britons, a higher PSH is associated with worse cognitive performance. These findings suggest that genetic predisposition to hypertension influences brain health in persons who have not yet developed dementia. Because information on genetic risk variants for elevated BP is available long before the development of hypertension, these results lay the foundation for further research focused on using genomic data for the early identification of high-risk middle-aged adults.

Suboptimal Sleep Duration is Associated with Poorer Neuroimaging Brain Health Profiles.

Background: Cardiovascular health optimization during middle age benefits brain health. The American Heart Association's Life's Simple 7 recently added sleep duration as a key determinant of cardiovascular health becoming the Life's Essential 8. We tested the hypothesis that suboptimal sleep duration is associated with poorer neuroimaging brain health profiles in asymptomatic middle-aged adults. Methods: We conducted a prospective MRI neuroimaging study in middle-aged persons without stroke, dementia, or multiple sclerosis enrolled in the UK Biobank. Self-reported sleep duration was categorized as short (<7 hours), optimal (7-<9 hours), or long (≥9 hours). Evaluated neuroimaging markers of brain health included white matter hyperintensities (presence and volume) and diffusion tensor imaging metrics (fractional anisotropy and mean diffusivity) evaluated in 48 distinct neuroanatomical regions. We used multivariable logistic and linear regression models, as appropriate, to test for association between sleep duration and neuroimaging markers of brain health. Results: We evaluated 39,502 middle-aged persons (mean age 55, 53% female). Of these, 28,712 (72.7%) had optimal, 8,422 (21.3%) short, and 2,368 (6%) long sleep. Compared to optimal sleep, short sleep was associated with higher risk (OR 1.11; 95% CI 1.05-1.17; P<0.001) and larger volume (beta=0.06, SE=0.01; P<0.001) of white matter hyperintensities, while long sleep was associated with higher volume (beta=0.04, SE=0.02; P=0.01) but not higher risk (P>0.05) of white matter hyperintensities. Short (beta=0.03, SE=0.01; P=0.004) and long sleep (beta=0.07, SE=0.02; P<0.001) were associated with worse fractional anisotropy, while only long sleep associated with worse mean diffusivity (beta=0.05, SE=0.02; P=0.005). Conclusions: Among middle-aged adults without clinically observed neurological disease, suboptimal sleep duration is associated with poorer neuroimaging brain health profiles. Because the evaluated neuroimaging markers precede stroke and dementia by several years, our findings support early interventions aimed at correcting this modifiable risk factor.

Screening for extranodal extension in HPV-associated oropharyngeal carcinoma: evaluation of a CT-based deep learning algorithm in patient data from a multicentre, randomised de-escalation trial.

BACKGROUND: Pretreatment identification of pathological extranodal extension (ENE) would guide therapy de-escalation strategies for in human papillomavirus (HPV)-associated oropharyngeal carcinoma but is diagnostically challenging. ECOG-ACRIN Cancer Research Group E3311 was a multicentre trial wherein patients with HPV-associated oropharyngeal carcinoma were treated surgically and assigned to a pathological risk-based adjuvant strategy of observation, radiation, or concurrent chemoradiation. Despite protocol exclusion of patients with overt radiographic ENE, more than 30% had pathological ENE and required postoperative chemoradiation. We aimed to evaluate a CT-based deep learning algorithm for prediction of ENE in E3311, a diagnostically challenging cohort wherein algorithm use would be impactful in guiding decision-making. METHODS: For this retrospective evaluation of deep learning algorithm performance, we obtained pretreatment CTs and corresponding surgical pathology reports from the multicentre, randomised de-escalation trial E3311. All enrolled patients on E3311 required pretreatment and diagnostic head and neck imaging; patients with radiographically overt ENE were excluded per study protocol. The lymph node with largest short-axis diameter and up to two additional nodes were segmented on each scan and annotated for ENE per pathology reports. Deep learning algorithm performance for ENE prediction was compared with four board-certified head and neck radiologists. The primary endpoint was the area under the curve (AUC) of the receiver operating characteristic. FINDINGS: From 178 collected scans, 313 nodes were annotated: 71 (23%) with ENE in general, 39 (13%) with ENE larger than 1 mm ENE. The deep learning algorithm AUC for ENE classification was 0·86 (95% CI 0·82-0·90), outperforming all readers (p<0·0001 for each). Among radiologists, there was high variability in specificity (43-86%) and sensitivity (45-96%) with poor inter-reader agreement (κ 0·32). Matching the algorithm specificity to that of the reader with highest AUC (R2, false positive rate 22%) yielded improved sensitivity to 75% (+ 13%). Setting the algorithm false positive rate to 30% yielded 90% sensitivity. The algorithm showed improved performance compared with radiologists for ENE larger than 1 mm (p<0·0001) and in nodes with short-axis diameter 1 cm or larger. INTERPRETATION: The deep learning algorithm outperformed experts in predicting pathological ENE on a challenging cohort of patients with HPV-associated oropharyngeal carcinoma from a randomised clinical trial. Deep learning algorithms should be evaluated prospectively as a treatment selection tool. FUNDING: ECOG-ACRIN Cancer Research Group and the National Cancer Institute of the US National Institutes of Health.

Clinical implementation of artificial intelligence in neuroradiology with development of a novel workflow-efficient picture archiving and communication system-based automated brain tumor segmentation and radiomic feature extraction.

Purpose: Personalized interpretation of medical images is critical for optimum patient care, but current tools available to physicians to perform quantitative analysis of patient's medical images in real time are significantly limited. In this work, we describe a novel platform within PACS for volumetric analysis of images and thus development of large expert annotated datasets in parallel with radiologist performing the reading that are critically needed for development of clinically meaningful AI algorithms. Specifically, we implemented a deep learning-based algorithm for automated brain tumor segmentation and radiomics extraction, and embedded it into PACS to accelerate a supervised, end-to- end workflow for image annotation and radiomic feature extraction. Materials and methods: An algorithm was trained to segment whole primary brain tumors on FLAIR images from multi-institutional glioma BraTS 2021 dataset. Algorithm was validated using internal dataset from Yale New Haven Health (YHHH) and compared (by Dice similarity coefficient [DSC]) to radiologist manual segmentation. A UNETR deep-learning was embedded into Visage 7 (Visage Imaging, Inc., San Diego, CA, United States) diagnostic workstation. The automatically segmented brain tumor was pliable for manual modification. PyRadiomics (Harvard Medical School, Boston, MA) was natively embedded into Visage 7 for feature extraction from the brain tumor segmentations. Results: UNETR brain tumor segmentation took on average 4 s and the median DSC was 86%, which is similar to published literature but lower than the RSNA ASNR MICCAI BRATS challenge 2021. Finally, extraction of 106 radiomic features within PACS took on average 5.8 ± 0.01 s. The extracted radiomic features did not vary over time of extraction or whether they were extracted within PACS or outside of PACS. The ability to perform segmentation and feature extraction before radiologist opens the study was made available in the workflow. Opening the study in PACS, allows the radiologists to verify the segmentation and thus annotate the study. Conclusion: Integration of image processing algorithms for tumor auto-segmentation and feature extraction into PACS allows curation of large datasets of annotated medical images and can accelerate translation of research into development of personalized medicine applications in the clinic. The ability to use familiar clinical tools to revise the AI segmentations and natively embedding the segmentation and radiomic feature extraction tools on the diagnostic workstation accelerates the process to generate ground-truth data.

Identifying clinically applicable machine learning algorithms for glioma segmentation: recent advances and discoveries.

Background: While there are innumerable machine learning (ML) research algorithms used for segmentation of gliomas, there is yet to be a US FDA cleared product. The aim of this study is to explore the systemic limitations of research algorithms that have prevented translation from concept to product by a review of the current research literature. Methods: We performed a systematic literature review on 4 databases. Of 11 727 articles, 58 articles met the inclusion criteria and were used for data extraction and screening using TRIPOD. Results: We found that while many articles were published on ML-based glioma segmentation and report high accuracy results, there were substantial limitations in the methods and results portions of the papers that result in difficulty reproducing the methods and translation into clinical practice. Conclusions: In addition, we identified that more than a third of the articles used the same publicly available BRaTS and TCIA datasets and are responsible for the majority of patient data on which ML algorithms were trained, which leads to limited generalizability and potential for overfitting and bias.

Outcomes after Thrombectomy for Minor Stroke: A Meta-Analysis.

PURPOSE: To determine the outcomes after mechanical thrombectomy (MT) versus medical management in patients with minor stroke symptomatology. METHODS: A meta-analysis was performed for studies reporting outcomes after MT, either as stand-alone therapy or with intravenous thrombolysis in patients with minor stroke and large-vessel occlusion. RESULTS: Fourteen studies with 2134 patients met the selection criteria and were included. Two studies compared immediate thrombectomy versus best medical management (with rescue thrombectomy) and the odds ratios of excellent outcomes, good outcomes, mortality and incidence of symptomatic intracranial hemorrhage (sICH) after immediate thrombectomy versus best medical management were 1.07 (95% confidence interval [CI] 0.93-1.22%), 1.15 (95% CI 1.05-1.25), 0.65 (95% CI 0.30-1.38), and 2.89 (95% CI 0.82-10.13), respectively. Among the 8 studies that compared MT outcomes versus medical management (without thrombectomy), odds ratios of excellent outcomes, good outcomes, mortality, and incidence of sICH after MT versus medical management were 0.98 (95% CI 0.89-1.07), 0.94 (95% CI 0.89-1.00), 1.61 (95% CI 1.08-2.41), and 2.59 (95% CI 1.35-4.96), respectively. Among all 14 studies, pooled proportions of excellent outcomes, good outcomes, mortality, and sICH after thrombectomy were 58.7%, 76.2%, 6.82%, and 3.23%, respectively. CONCLUSIONS: Our study shows significant selection bias and heterogeneity in the literature with differences in baseline characteristics (age, stroke severity, prestroke modified Rankin Scale score, side of infarct, vessel and site of occlusion, use of intravenous thrombolysis, criteria for clinical deterioration, and selection bias for rescue MT and rates of reperfusion), emphasizing the need for a randomized controlled trial.

Cost-effectiveness of endovascular thrombectomy in patients with acute stroke and M2 occlusion.

BACKGROUND: The cost-effectiveness of endovascular thrombectomy (EVT) in patients with acute ischemic stroke due to M2 branch occlusion remains uncertain. OBJECTIVE: To evaluate the cost-effectiveness of EVT compared with medical management in patients with acute stroke presenting with M2 occlusion using a decision-analytic model. METHODS: A decision-analytic study was performed with Markov modeling to estimate the lifetime quality-adjusted life years and associated costs of EVT-treated patients compared with no-EVT/medical management. The study was performed over a lifetime horizon with a societal perspective in the Unites States setting. Base case, one-way, two-way, and probabilistic sensitivity analyses were performed. RESULTS: EVT was the long-term cost-effective strategy in 93.37% of the iterations in the probabilistic sensitivity analysis, and resulted in difference in health benefit of 1.66 QALYs in the 65-year-old age groups, equivalent to 606 days in perfect health. Varying the outcomes after both strategies shows that EVT was more cost-effective when the probability of good outcome after EVT was only 4-6% higher relative to medical management in clinically likely scenarios. EVT remained cost-effective even when its cost exceeded US$200 000 (threshold was US$209 111). EVT was even more cost-effective for 55-year-olds than for 65-year-old patients. CONCLUSION: Our study suggests that EVT is cost-effective for treatment of acute M2 branch occlusions. Faster and improved reperfusion techniques would increase the relative cost-effectiveness of EVT even further in these patients.

Cost-effectiveness of endovascular thrombectomy in patients with low Alberta Stroke Program Early CT Scores (< 6) at presentation.

OBJECTIVE: The utility of endovascular thrombectomy (EVT) in patients with acute ischemic stroke, large vessel occlusion (LVO), and low Alberta Stroke Program Early CT Scores (ASPECTS) remains uncertain. The objective of this study was to determine the health outcomes and cost-effectiveness of EVT versus medical management in patients with ASPECTS < 6. METHODS: A decision-analytical study was performed with Markov modeling to estimate the lifetime quality-adjusted life-years (QALYs) and associated costs of EVT-treated patients compared to medical management. The study was performed over a lifetime horizon with a societal perspective in the US setting. RESULTS: The incremental cost-effectiveness ratios were $412,411/QALY and $1,022,985/QALY for 55- and 65-year-old groups in the short-term model. EVT was the long-term cost-effective strategy in 96.16% of the iterations and resulted in differences in health benefit of 2.21 QALYs and 0.79 QALYs in the 55- and 65-year-old age groups, respectively, equivalent to 807 days and 288 days in perfect health. EVT remained the more cost-effective strategy when the probability of good outcome with EVT was above 16.8% or as long as the good outcome associated with the procedure was at least 1.6% higher in absolute value than that of medical management. EVT remained cost-effective even when its cost exceeded $100,000 (threshold was $108,036). Although the cost-effectiveness decreased with age, EVT was cost-effective for 75-year-old patients as well. CONCLUSIONS: This study suggests that EVT is the more cost-effective approach compared to medical management in patients with ASPECTS < 6 in the long term (lifetime horizon), considering the poor outcomes and significant disability associated with nonreperfusion.

Vessel wall magnetic resonance imaging in intracranial aneurysms: Principles and emerging clinical applications.

Intracranial high-resolution vessel wall magnetic resonance imaging is an imaging paradigm that complements conventional imaging modalities used in the evaluation of neurovascular pathology. This review focuses on the emerging utility of vessel wall magnetic resonance imaging in the characterization of intracranial aneurysms. We first discuss the technical principles of vessel wall magnetic resonance imaging highlighting methods to determine aneurysm wall enhancement and how to avoid common interpretive pitfalls. We then review its clinical application in the characterization of ruptured and unruptured intracranial aneurysms, in particular, the emergence of aneurysm wall enhancement as a biomarker of aneurysm instability. We offer our perspective from a high-volume neurovascular center where vessel wall magnetic resonance imaging is in routine clinical use.

Multi-Institutional Validation of Deep Learning for Pretreatment Identification of Extranodal Extension in Head and Neck Squamous Cell Carcinoma.

PURPOSE: Extranodal extension (ENE) is a well-established poor prognosticator and an indication for adjuvant treatment escalation in patients with head and neck squamous cell carcinoma (HNSCC). Identification of ENE on pretreatment imaging represents a diagnostic challenge that limits its clinical utility. We previously developed a deep learning algorithm that identifies ENE on pretreatment computed tomography (CT) imaging in patients with HNSCC. We sought to validate our algorithm performance for patients from a diverse set of institutions and compare its diagnostic ability to that of expert diagnosticians. METHODS: We obtained preoperative, contrast-enhanced CT scans and corresponding pathology results from two external data sets of patients with HNSCC: an external institution and The Cancer Genome Atlas (TCGA) HNSCC imaging data. Lymph nodes were segmented and annotated as ENE-positive or ENE-negative on the basis of pathologic confirmation. Deep learning algorithm performance was evaluated and compared directly to two board-certified neuroradiologists. RESULTS: A total of 200 lymph nodes were examined in the external validation data sets. For lymph nodes from the external institution, the algorithm achieved an area under the receiver operating characteristic curve (AUC) of 0.84 (83.1% accuracy), outperforming radiologists' AUCs of 0.70 and 0.71 (P = .02 and P = .01). Similarly, for lymph nodes from the TCGA, the algorithm achieved an AUC of 0.90 (88.6% accuracy), outperforming radiologist AUCs of 0.60 and 0.82 (P < .0001 and P = .16). Radiologist diagnostic accuracy improved when receiving deep learning assistance. CONCLUSION: Deep learning successfully identified ENE on pretreatment imaging across multiple institutions, exceeding the diagnostic ability of radiologists with specialized head and neck experience. Our findings suggest that deep learning has utility in the identification of ENE in patients with HNSCC and has the potential to be integrated into clinical decision making.