Stroke subtype classification by geometrical descriptors of lesion shape

Deep Learning-Based Automatic Classification of Ischemic Stroke Subtype Using Diffusion-Weighted Images

BACKGROUND AND PURPOSE

Accurate classification of ischemic stroke subtype is important for effective secondary prevention of stroke. We used diffusion-weighted image (DWI) and atrial fibrillation (AF) data to train a deep learning algorithm to classify stroke subtype.

METHODS

Model development was done in 2,988 patients with ischemic stroke from three centers by using U-net for infarct segmentation and EfficientNetV2 for subtype classification. Experienced neurologists (n=5) determined subtypes for external test datasets, while establishing a consensus for clinical trial datasets. Automatically segmented infarcts were fed into the model (DWI-only algorithm). Subsequently, another model was trained, with AF included as a categorical variable (DWI+AF algorithm). These models were tested: (1) internally against the opinion of the labeling experts, (2) against fresh external DWI data, and (3) against clinical trial dataset.

RESULTS

In the training-and-validation datasets, the mean (±standard deviation) age was 68.0±12.5 (61.1% male). In internal testing, compared with the experts, the DWI-only and the DWI+AF algorithms respectively achieved moderate (65.3%) and near-strong (79.1%) agreement. In external testing, both algorithms again showed good agreements (59.3%-60.7% and 73.7%-74.0%, respectively). In the clinical trial dataset, compared with the expert consensus, percentage agreements and Cohen's kappa were respectively 58.1% and 0.34 for the DWI-only vs. 72.9% and 0.57 for the DWI+AF algorithms. The corresponding values between experts were comparable (76.0% and 0.61) to the DWI+AF algorithm.

CONCLUSION

Our model trained on a large dataset of DWI (both with or without AF information) was able to classify ischemic stroke subtypes comparable to a consensus of stroke experts.

Predicting Hypoperfusion Lesion and Target Mismatch in Stroke from Diffusion-weighted MRI Using Deep Learning

Background Perfusion imaging is important to identify a target mismatch in stroke but requires contrast agents and postprocessing software. Purpose To use a deep learning model to predict the hypoperfusion lesion in stroke and identify patients with a target mismatch profile from diffusion-weighted imaging (DWI) and clinical information alone, using perfusion MRI as the reference standard. Materials and Methods Imaging data sets of patients with acute ischemic stroke with baseline perfusion MRI and DWI were retrospectively reviewed from multicenter data available from 2008 to 2019 (Imaging Collaterals in Acute Stroke, Diffusion and Perfusion Imaging Evaluation for Understanding Stroke Evolution 2, and University of California, Los Angeles stroke registry). For perfusion MRI, rapid processing of perfusion and diffusion software automatically segmented the hypoperfusion lesion (time to maximum, ≥6 seconds) and ischemic core (apparent diffusion coefficient [ADC], ≤620 × 10-6 mm2/sec). A three-dimensional U-Net deep learning model was trained using baseline DWI, ADC, National Institutes of Health Stroke Scale score, and stroke symptom sidedness as inputs, with the union of hypoperfusion and ischemic core segmentation serving as the ground truth. Model performance was evaluated using the Dice score coefficient (DSC). Target mismatch classification based on the model was compared with that of the clinical-DWI mismatch approach defined by the DAWN trial by using the McNemar test. Results Overall, 413 patients (mean age, 67 years ± 15 [SD]; 207 men) were included for model development and primary analysis using fivefold cross-validation (247, 83, and 83 patients in the training, validation, and test sets, respectively, for each fold). The model predicted the hypoperfusion lesion with a median DSC of 0.61 (IQR, 0.45-0.71). The model identified patients with target mismatch with a sensitivity of 90% (254 of 283; 95% CI: 86, 93) and specificity of 77% (100 of 130; 95% CI: 69, 83) compared with the clinical-DWI mismatch sensitivity of 50% (140 of 281; 95% CI: 44, 56) and specificity of 89% (116 of 130; 95% CI: 83, 94) (P < .001 for all). Conclusion A three-dimensional U-Net deep learning model predicted the hypoperfusion lesion from diffusion-weighted imaging (DWI) and clinical information and identified patients with a target mismatch profile with higher sensitivity than the clinical-DWI mismatch approach. ClinicalTrials.gov registration nos. NCT02225730, NCT01349946, NCT02586415 © RSNA, 2022 Supplemental material is available for this article. See also the editorial by Kallmes and Rabinstein in this issue.

Predicting final ischemic stroke lesions from initial diffusion-weighted images using a deep neural network

BACKGROUND

For prognosis of stroke, measurement of the diffusion-perfusion mismatch is a common practice for estimating tissue at risk of infarction in the absence of timely reperfusion. However, perfusion-weighted imaging (PWI) adds time and expense to the acute stroke imaging workup. We explored whether a deep convolutional neural network (DCNN) model trained with diffusion-weighted imaging obtained at admission could predict final infarct volume and location in acute stroke patients.

METHODS

In 445 patients, we trained and validated an attention-gated (AG) DCNN to predict final infarcts as delineated on follow-up studies obtained 3 to 7 days after stroke. The input channels consisted of MR diffusion-weighted imaging (DWI), apparent diffusion coefficients (ADC) maps, and thresholded ADC maps with values less than 620 × 10-6 mm2/s, while the output was a voxel-by-voxel probability map of tissue infarction. We evaluated performance of the model using the area under the receiver-operator characteristic curve (AUC), the Dice similarity coefficient (DSC), absolute lesion volume error, and the concordance correlation coefficient (ρc) of the predicted and true infarct volumes.

RESULTS

The model obtained a median AUC of 0.91 (IQR: 0.84-0.96). After thresholding at an infarction probability of 0.5, the median sensitivity and specificity were 0.60 (IQR: 0.16-0.84) and 0.97 (IQR: 0.93-0.99), respectively, while the median DSC and absolute volume error were 0.50 (IQR: 0.17-0.66) and 27 ml (IQR: 7-60 ml), respectively. The model's predicted lesion volumes showed high correlation with ground truth volumes (ρc = 0.73, p < 0.01).

CONCLUSION

An AG-DCNN using diffusion information alone upon admission was able to predict infarct volumes at 3-7 days after stroke onset with comparable accuracy to models that consider both DWI and PWI. This may enable treatment decisions to be made with shorter stroke imaging protocols.

Association of stroke lesion shape with newly detected atrial fibrillation – Results from the MonDAFIS study

Paroxysmal Atrial fibrillation (AF) is often clinically silent and may be missed by the usual diagnostic workup after ischemic stroke. We aimed to determine whether shape characteristics of ischemic stroke lesions can be used to predict AF in stroke patients without known AF at baseline. Lesion shape quantification on brain MRI was performed in selected patients from the intervention arm of the Impact of standardized MONitoring for Detection of Atrial Fibrillation in Ischemic Stroke (MonDAFIS) study, which included patients with ischemic stroke or TIA without prior AF. Multiple morphologic parameters were calculated based on lesion segmentation in acute brain MRI data. Multivariate logistic models were used to test the association of lesion morphology, clinical parameters, and AF. A stepwise elimination regression was conducted to identify the most important variables. A total of 755 patients were included. Patients with AF detected within 2 years after stroke (n = 86) had a larger overall oriented bounding box (OBB) volume (p = 0.003) and a higher number of brain lesion components (p = 0.008) than patients without AF. In the multivariate model, OBB volume (OR 1.72, 95%CI 1.29–2.35, p < 0.001), age (OR 2.13, 95%CI 1.52–3.06, p < 0.001), and female sex (OR 2.45, 95%CI 1.41–4.31, p = 0.002) were independently associated with detected AF. Ischemic lesions in patients with detected AF after stroke presented with a more dispersed infarct pattern and a higher number of lesion components. Together with clinical characteristics, these lesion shape characteristics may help in guiding prolonged cardiac monitoring after stroke.

Phybrata Sensors and Machine Learning for Enhanced Neurophysiological Diagnosis and Treatment

Concussion injuries remain a significant public health challenge. A significant unmet clinical need remains for tools that allow related physiological impairments and longer-term health risks to be identified earlier, better quantified, and more easily monitored over time. We address this challenge by combining a head-mounted wearable inertial motion unit (IMU)-based physiological vibration acceleration (“phybrata”) sensor and several candidate machine learning (ML) models. The performance of this solution is assessed for both binary classification of concussion patients and multiclass predictions of specific concussion-related neurophysiological impairments. Results are compared with previously reported approaches to ML-based concussion diagnostics. Using phybrata data from a previously reported concussion study population, four different machine learning models (Support Vector Machine, Random Forest Classifier, Extreme Gradient Boost, and Convolutional Neural Network) are first investigated for binary classification of the test population as healthy vs. concussion (Use Case 1). Results are compared for two different data preprocessing pipelines, Time-Series Averaging (TSA) and Non-Time-Series Feature Extraction (NTS). Next, the three best-performing NTS models are compared in terms of their multiclass prediction performance for specific concussion-related impairments: vestibular, neurological, both (Use Case 2). For Use Case 1, the NTS model approach outperformed the TSA approach, with the two best algorithms achieving an F1 score of 0.94. For Use Case 2, the NTS Random Forest model achieved the best performance in the testing set, with an F1 score of 0.90, and identified a wider range of relevant phybrata signal features that contributed to impairment classification compared with manual feature inspection and statistical data analysis. The overall classification performance achieved in the present work exceeds previously reported approaches to ML-based concussion diagnostics using other data sources and ML models. This study also demonstrates the first combination of a wearable IMU-based sensor and ML model that enables both binary classification of concussion patients and multiclass predictions of specific concussion-related neurophysiological impairments.

Physiological Vibration Acceleration (Phybrata) Sensor Assessment of Multi-System Physiological Impairments and Sensory Reweighting Following Concussion

Objective

To assess the utility of a head-mounted wearable inertial motion unit (IMU)-based physiological vibration acceleration (“phybrata”) sensor to support the clinical diagnosis of concussion, classify and quantify specific concussion-induced physiological system impairments and sensory reweighting, and track individual patient recovery trajectories.

Methods

Data were analyzed from 175 patients over a 12-month period at three clinical sites. Comprehensive clinical concussion assessments were first completed for all patients, followed by testing with the phybrata sensor. Phybrata time series data and spatial scatter plots, eyes open (Eo) and eyes closed (Ec) phybrata powers, average power (Eo+Ec)/2, Ec/Eo phybrata power ratio, time-resolved phybrata spectral density (TRPSD) distributions, and receiver operating characteristic (ROC) curves are compared for individuals with no objective impairments and those clinically diagnosed with concussions and accompanying vestibular impairment, other neurological impairment, or both vestibular and neurological impairments. Finally, pre- and post-injury phybrata case report results are presented for a participant who was diagnosed with a concussion and subsequently monitored during treatment, rehabilitation, and return-to-activity clearance.

Results

Phybrata data demonstrate distinct features and patterns for individuals with no discernable clinical impairments, diagnosed vestibular pathology, and diagnosed neurological pathology. ROC curves indicate that the average power (Eo+Ec)/2 may be utilized to support clinical diagnosis of concussion, while Eo and Ec/Eo may be utilized as independent measures to confirm accompanying neurological and vestibular impairments, respectively. All 3 measures demonstrate area under the curve (AUC), sensitivity, and specificity above 90% for their respective diagnoses. Phybrata spectral analyses demonstrate utility for quantifying the severity of concussion-induced physiological impairments, sensory reweighting, and subsequent monitoring of improvements throughout treatment and rehabilitation.

Conclusion

Phybrata testing assists with objective concussion diagnosis and provides an important adjunct to standard concussion assessment tools by objectively ascertaining neurological and vestibular impairments, guiding targeted rehabilitation strategies, monitoring recovery, and assisting with return-to-sport/work/learn decision-making.

A novel approach to treatment of thromboembolic stroke in mice: Redirecting neutrophils toward a peripherally implanted CXCL1-soaked sponge

Neutrophils are considered key participants in post-ischemic stroke inflammation. They are the first white blood cells to arrive in ischemic brain and their presence in the brain tissue positively correlates with post-ischemic injury severity. CXCL1 is a neutrophil attractant chemokine and the present study evaluates whether redirecting neutrophil migration using a peripherally implanted CXCL1-soaked sponge can reduce brain inflammation and improve outcomes in a novel mouse model of thromboembolic (TE) stroke. TE stroke was induced by injection of a platelet-rich microemboli suspension into the internal carotid artery of adult C57BL/6 male mice. The model induced neuroinflammation that was associated with increases in multiple brain and serum cytokines/chemokines at the mRNA and protein levels, including very marked increases in CXCL1. In other groups of animals, an absorbable sterile hemostatic sponge, previously immersed in either saline (0.9%NaCl) or CXCL1, was implanted into subcutaneous pockets formed in the inguinal region on the left and right side following stroke surgery. Mice implanted with the sponge soaked with CXCL1 had significantly reduced neuroinflammation and infarct size after TE stroke compared to mice implanted with the sponge soaked with 0.9%NaCl. There was also reduced mortality and improved neurological deficits in the TE stroke + CXCL1 sponge group compared to the TE stroke +0.9%NaCl sponge group. In conclusion: redirecting bloodstream leukocytes toward a peripherally-implanted neutrophil chemokine CXCL1-soaked sponge improves outcomes in a novel mouse model of thromboembolic stroke. The present findings suggest a novel therapeutic strategy for patients with acute stroke.

Three cooperative mechanisms required for recovery after brain damage

Stroke is one of the main causes of human disabilities. Experimental observations indicate that several mechanisms are activated during the recovery of functional activity after a stroke. Here we unveil how the brain recovers by explaining the role played by three mechanisms: Plastic adaptation, hyperexcitability and synaptogenesis. We consider two different damages in a neural network: A diffuse damage that simply causes the reduction of the effective system size and a localized damage, a stroke, that strongly alters the spontaneous activity of the system. Recovery mechanisms observed experimentally are implemented both separately and in a combined way. Interestingly, each mechanism contributes to the recovery to a limited extent. Only the combined application of all three together is able to recover the spontaneous activity of the undamaged system. This explains why the brain triggers independent mechanisms, whose cooperation is the fundamental ingredient for the system’s recovery.

Noncardioembolic Stroke in Patients with Atrial Fibrillation

Atrial fibrillation (AF) could be a coincidental finding in certain patients with ischemic stroke and increased burden of underlying cardiovascular disease. Concomitant large-vessel atheromatosis and cerebral small vessel disease may be the actual cause of stroke, and distinguishing between different pathophysiologic mechanisms could impose substantial diagnostic difficulties. Despite routine use of oral anticoagulants (OACs) in patients with AF based on their risk for embolism (ie, CHA2DS2-Vasc score), antithrombotic agents may exert differential effects depending on stroke etiology and stroke subtyping should be evaluated as an additional component of risk stratification that could facilitate optimal management. In the present study, we summarize the evidence on noncardioembolic (non-CE) stroke and treatment approaches based on different stroke subtypes in patients with AF. In particular, approximately one-third of patients with AF seem to suffer a non-CE stroke. Within this category, 11% to 24% of patients present high-grade carotid stenosis and 9% to 16% of ischemic strokes are classified as lacunar. In terms of secondary prevention, the effectiveness of OACs in preventing non-CE stroke has been disputed. Additional large-scale prospective studies are warranted to assess the pathophysiologic stroke mechanisms in patients with AF and compare the differential efficacy of antithrombotic treatment strategies in non-CE ischemic syndromes.