Fundus Retinal imaging is an easy-to-acquire modality typically used for monitoring eye health. Current evidence indicates that the retina, and its vasculature in particular, is associated with other disease processes making it an ideal candidate for biomarker discovery. The development of these biomarkers has typically relied on predefined measurements, which makes the development process slow. Recently, representation learning algorithms such as general purpose convolutional neural networks or vasculature embeddings have been proposed as an approach to learn imaging biomarkers directly from the data, hence greatly speeding up their discovery. In this work, we compare and contrast different state-of-the-art retina biomarker discovery methods to identify signs of past stroke in the retinas of a curated patient cohort of 2,472 subjects from the UK Biobank dataset. We investigate two convolutional neural networks previously used in retina biomarker discovery and directly trained on the stroke outcome, and an extension of the vasculature embedding approach which infers its feature representation from the vasculature and combines the information of retinal images from both eyes.In our experiments, we show that the pipeline based on vasculature embeddings has comparable or better performance than other methods with a much more compact feature representation and ease of training.Clinical Relevance-This study compares and contrasts three retinal biomarker discovery strategies, using a curated dataset of subject evidence, for the analysis of the retina as a proxy in the assessment of clinical outcomes, such as stroke risk.
Vessel segmentation in fundus images permits understanding retinal diseases and computing image-based biomarkers. However, manual vessel segmentation is a time-consuming process. Optical coherence tomography angiography (OCT-A) allows direct, non-invasive estimation of retinal vessels. Unfortunately, compared to fundus images, OCT-A cameras are more expensive, less portable, and have a reduced field of view. We present an automated strategy relying on generative adversarial networks to create vascular maps from fundus images without training using manual vessel segmentation maps. Further post-processing used for standard en face OCT-A allows obtaining a vessel segmentation map. We compare our approach to state-of-the-art vessel segmentation algorithms trained on manual vessel segmentation maps and vessel segmentations derived from OCT-A. We evaluate them from an automatic vascular segmentation perspective and as vessel density estimators, i.e., the most common imaging biomarker for OCT-A used in studies. Using OCT-A as a training target over manual vessel delineations yields improved vascular maps for the optic disc area and compares to the best-performing vessel segmentation algorithm in the macular region. This technique could reduce the cost and effort incurred when training vessel segmentation algorithms. To incentivize research in this field, we will make the dataset publicly available to the scientific community.
Introduction: Early neurological deterioration (END) in patients with large vessel occlusion (LVO) stroke with low NIHSS treated with medical management (MM) has been well described. However, the effect of END on 90-day disability outcomes relative to patients presenting with more severe stroke, remains undetermined. Methods: From our multi-center prospective registry across 4 comprehensive stroke centers, we identified patients with LVO (basilar, ICA, M1/2, or P1 identified on CTA or MRA) from January 2018 to June 2020. Low NIHSS was defined as 0-5, and END as a worsening of > 3 points. To determine the effect of END on outcomes, we used propensity score to match patients who presented with low NIHSS who suffered END by age and worsened NIHSS (i.e. subsequent, greater NIHSS) to patients who presented initially with LVO and comparable NIHSS. Results: Among 348 patients with LVO acute ischemic stroke, median age was 67 [IQR 59-76], 46.6% were female. 58 patients (17%) of the cohort had low NIHSS. Compared to the higher NIHSS group, low NIHSS group had less EVT performed (39.7% vs 78.3%, p=<0.001), greater ASPECTS (9 vs 8, p=0.001), fewer M1 occlusions (27.6% vs 51.7%, p<0.001), and lower prevalence of risk factors, such as hypertension, hyperlipidemia, and congestive heart failure (50.0% vs 71.4, p=0.001; 20.7% vs 36.9%, p=0.018; and 0.0% vs 9.3%, p=0.016, respectively). Overall 90-day disability outcomes (mRS 0-2) were better in the low NIHSS group (58.6% vs 44.8%, p=0.055) although it did not reach statistical significance. On the other hand, 41.4% of low NIHSS group did not achieve functional independence. Of 35 patients with initial MM in the low NIHSS group, 7 (20%) suffered END. The median increase in NIHSS was 7 (range 4-15), and only 1 patient achieved 90-day mRS 0-2. In a propensity matching analysis, LVO patients with initially low NIHSS patients followed by END had significantly worse outcomes than the patients who initially presented with comparable NIHSS (Coef=26, p=<0.001). Conclusions: In patients with low NIHSS and LVO, those treated initially with MM who suffer END have worse clinical outcomes when matched to patients with comparable NIHSS post-worsening. These finding suggest that the delaying EVT in this population may lead to worsened outcomes.
Background: Prehospital automated large vessel occlusion (LVO) detection in Mobile Stroke Units (MSUs) could accelerate identification and treatment of patients with LVO acute ischemic stroke. Here, we evaluate the performance of a machine learning (ML) model on CT angiograms (CTAs) obtained from 2 MSUs to detect LVO. Methods: Patients evaluated on MSUs in Houston and Los Angeles with out-of-hospital CTAs were identified. Anterior circulation LVO was defined as an occlusion of the intracranial internal carotid artery, middle cerebral artery (M1 or M2), or anterior cerebral artery vessels and determined by an expert human reader. A ML model to detect LVO was trained and tested on independent data sets consisting of in-hospital CTAs and then tested on MSU CTA images. Model performance was determined using area under the receiver-operator curve statistics. Results: Among 68 patients with out-of-hospital MSU CTAs, 40% had an LVO. The most common occlusion location was the middle cerebral artery M1 segment (59%), followed by the internal carotid artery (30%), and middle cerebral artery M2 (11%). Median time from last known well to CTA imaging was 88.0 (interquartile range, 59.5–196.0) minutes. After training on 870 in-hospital CTAs, the ML model performed well in identifying LVO in a separate in-hospital data set of 441 images with area under receiver-operator curve of 0.84 (95% CI, 0.80–0.87). ML algorithm analysis time was under 1 minute. The performance of the ML model on the MSU CTA images was comparable with area under receiver-operator curve 0.80 (95% CI, 0.71–0.89). There was no significant difference in performance between the Houston and Los Angeles MSU CTA cohorts. Conclusions: In this study of patients evaluated on MSUs in 2 cities, a ML algorithm was able to accurately and rapidly detect LVO using prehospital CTA acquisitions.
Background: Intracerebral hemorrhage (ICH) constitutes upto 40% mortality in first 30 days. Early identification of predictors of hematoma expansion (HE) may improve efforts to prevent its occurrence and improve clinical outcome. Methods: We identified patients with ICH and follow-up imaging. HE was defined as a combination of absolute volume increase of 6cc, new IVH, or proportional increase of 33% in our dataset on 72h follow up scan. Presence of IVH was also included in hematoma expansion. We evaluated the predictive ability of 3 machine learning classifiers, Random Forest, Support Vector Machine (with RBF kernel) and Logistic regression (with L1 regularization). The evaluation was done using a K-fold stratified cross validation to avoid overfitting. K was selected to be the number of subjects with HE. The features employed by classifiers were entirely based on the baseline imaging: Hematoma volume, Systolic BP, Diastolic BP, Black hole signs, Island signs, Blend signs, Fluid level, Swirl signs, Spot signs. Results: Our dataset comprised of 91 patients (n=21 HE, n=70 no HE). According to the area under the ROC (AUC), the two top performing classifiers were Support Vector Machine (AUC=0.66 CI 0.50-0.79) and Logistic Regression (AUC=0.64 CI 0.49-0.80). The statistical significance of the prediction is confirmed by the Mann-Whitney U test, p=0.01 and p=0.04 respectively. Random Forest did not reach statistical significance. Finally, we evaluated what were the highest and lowest weighted features across the cross-validation with Logistic Regression. The 3 top features were: presence of black hole and island signs and the systolic blood pressure. The 3 least useful features were: presence of spot and swirl signs and hematoma volume. Conclusion: Using our cohort, we developed a machine learning algorithm that predicts hematoma expansion using imaging features and blood pressure. MBL provided better sensitivity of these imaging markers compared with previous studies.
To derive, test and externally validate LVO incidence measures using widely available data points, clinical findings, and imaging features on non-contrast head CT.
Scarcity of labels for medical images is a significant barrier for training representation learning approaches based on deep neural networks. This limitation is also present when using imaging data collected during routine clinical care stored in picture archiving communication systems (PACS), as these data rarely have attached the high-quality labels required for medical image computing tasks. However, medical images extracted from PACS are commonly coupled with descriptive radiology reports that contain significant information and could be leveraged to pre-train imaging models, which could serve as starting points for further task-specific fine-tuning. In this work, we perform a head-to-head comparison of three different self-supervised strategies to pre-train the same imaging model on 3D brain computed tomography angiogram (CTA) images, with large vessel occlusion (LVO) detection as the downstream task. These strategies evaluate two natural language processing (NLP) approaches, one to extract 100 explicit radiology concepts (Rad-SpatialNet) and the other to create general-purpose radiology reports embeddings (DistilBERT). In addition, we experiment with learning radiology concepts directly or by using a recent self-supervised learning approach (CLIP) that learns by ranking the distance between language and image vector embeddings. The LVO detection task was selected because it requires 3D imaging data, is clinically important, and requires the algorithm to learn outputs not explicitly stated in the radiology report. Pre-training was performed on an unlabeled dataset containing 1,542 3D CTA - reports pairs. The downstream task was tested on a labeled dataset of 402 subjects for LVO. We find that the pre-training performed with CLIP-based strategies improve the performance of the imaging model to detect LVO compared to a model trained only on the labeled data. The best performance was achieved by pre-training using the explicit radiology concepts and CLIP strategy.
The optimal endovascular stroke therapy (EVT) care delivery structure is unknown. Here, we present our experience in creating an integrated stroke system (ISS) to expand EVT availability throughout our region while maintaining hospital and physician quality standards.We identified all consecutive patients with large vessel occlusion acute ischemic stroke treated with EVT from January 2014 to February 2019 in our health care system. In October 2017, we implemented the ISS, in which 3 additional hospitals (4 total) became EVT-performing hospitals (EPHs) and physicians were rotated between all centers. The cohort was divided by time into pre-ISS and post-ISS, and the primary outcome was time from stroke onset to EPH arrival. Secondary outcomes included hospital and procedural quality metrics. We performed an external validation using data from the Southeast Texas Regional Advisory Council.Among 513 patients with large vessel occlusion acute ischemic stroke treated with EVT, 58% were treated pre-ISS and 43% post-ISS. Over the study period, EVT procedural volume increased overall but remained relatively low at the 3 new EPHs (<70 EVT/y). After ISS, the proportion of patients who underwent interhospital transfer decreased (46% versus 37%; P<0.05). In adjusted quantile regression, ISS implementation resulted in a reduction of time from stroke onset to EPH arrival by 40 minutes (P<0.01) and onset to groin puncture by 29 minutes (P<0.05). Rates of postprocedural hemorrhage, modified Thrombolysis in Cerebral Infarction (TICI) 2b/3, and 90-day modified Rankin Scale were comparable at the higher and lower volume EPHs. The improvement in onset-to-arrival time was not reflective of overall improvement in secular trends in regional prehospital care.In our system, increasing EVT availability decreased time from stroke onset to EPH arrival. The ISS provides a framework to maintain quality in lower volume hospitals.
Introduction: The optimal blood pressure (BP) in patients with large vessel occlusion (LVO) acute ischemic stroke (AIS). Here, we explore the relationship between CT perfusion (CTP) predicted infarct volumes and those seen on follow up MRI, to examine whether infarct growth is related to presentation BP. Methods: From our prospectively collected multi-institutional registry, we identified patients with LVO AIS seen at 4 comprehensive stroke centers from January 2018 to March 2021. Patients were included if they contained anterior circulation LVO (defined as occlusion of the intracranial ICA, MCA or ACA) defined by CTA, included if they underwent CTP with RAPID (IschemiaView) post-processing at the time of presentation, and had final infarct volume (FIV) imaging with MRI 48-72 hours later. Infarct growth was defined as increase in infarct volume of at least 10 mL from CTP-RAPID core volume prediction to FIV. The primary outcome was the effect of presentation mean arterial blood pressure (MAP) on likelihood of infarct growth. Results: Among 329 patients that met inclusion criteria, median age was 68 [IQR 58-70], median NIHSS was 15 [IQR 10-20] and 49% were female. Median ASPECTS was 8 [6-9], median CTP-RAPID core was 6 mL [0-36 mL] and median FIV was 19 mL [5-57 mL]. Median arrival systolic BP was 153 mmHg [132-174 mmHg], diastolic was 83 mmHg [73-92 mmHg] and MAP was 105 mmHg [95- 118]. 161 (49%) of patients presented in the early time window. Infarct growth of at least 10 mL was seen in 23 (7%) of patients. FIV correlated with CTP-RAPID core more clearly for patients in the early vs. late window (R=0.77 vs 0.34, early vs. late window). In the subset of patients presenting in the early time window, those who underwent infarct growth of at least 10 mL were more likely to present with greater MAPs (mean MAP 118 vs. 106, infarct growth vs. no growth, p<0.05). This relationship was not seen in patients in the late window. Conclusions: Infarct expansion from presentation to FIV was seen in a substantial proportion of patients and associated with greater MAPs in the early but not late window. There are multiple possible explanations for this finding, including increased CTP inaccuracies in the early time window, and potentially with elevated BP. Further work on the optimal BP in LVO AIS is needed.
