Automatic Segmentation in Acute Ischemic Stroke: Prognostic Significance of Topological Stroke Volumes on Stroke Outcome.

BACKGROUND: Stroke infarct volume predicts patient disability and has utility for clinical trial outcomes. Accurate infarct volume measurement requires manual segmentation of stroke boundaries in diffusion-weighted magnetic resonance imaging scans which is time-consuming and subject to variability. Automatic infarct segmentation should be robust to rotation and reflection; however, prior work has not encoded this property into deep learning architecture. Here, we use rotation-reflection equivariance and train a deep learning model to segment stroke volumes in a large cohort of well-characterized patients with acute ischemic stroke in different vascular territories. METHODS: In this retrospective study, patients were selected from a stroke registry at Houston Methodist Hospital. Eight hundred seventy-five patients with acute ischemic stroke in any brain area who had magnetic resonance imaging with diffusion-weighted imaging were included for analysis and split 80/20 for training/testing. Infarct volumes were manually segmented by consensus of 3 independent clinical experts and cross-referenced against radiology reports. A rotation-reflection equivariant model was developed based on U-Net and grouped convolutions. Segmentation performance was evaluated using Dice score, precision, and recall. Ninety-day modified Rankin Scale outcome prediction was also evaluated using clinical variables and segmented stroke volumes in different brain regions. RESULTS: Segmentation model Dice scores are 0.88 (95% CI, 0.87-0.89; training) and 0.85 (0.82-0.88; testing). The modified Rankin Scale outcome prediction AUC using stroke volume in 30 refined brain regions based upon modified Rankin Scale-relevance areas adjusted for clinical variables was 0.80 (0.76-0.83) with an accuracy of 0.75 (0.72-0.78). CONCLUSIONS: We trained a deep learning model with encoded rotation-reflection equivariance to segment acute ischemic stroke lesions in diffusion- weighted imaging using a large data set from the Houston Methodist stroke center. The model achieved competitive performance in 175 well-balanced hold-out testing cases that include strokes from different vascular territories. Furthermore, the location specific stroke volume segmentations from the deep learning model combined with clinical factors demonstrated high AUC and accuracy for 90-day modified Rankin Scale in an outcome prediction model.

Reliability of the National Institutes of Health (NIH) Stroke Scale Between Emergency Room and Neurology Physicians for Initial Stroke Severity Scoring.

INTRODUCTION: In patients with acute ischemic stroke (AIS), the National Institutes of Health Stroke Scale (NIHSS) is essential to establishing a patient's initial stroke severity. While previous research has validated NIHSS scoring reliability between neurologists and other clinicians, it has not specifically evaluated NIHSS scoring reliability between emergency room (ER) and neurology physicians within the same clinical scenario and timeframe in a large cohort of patients. This study specifically addresses the key question: does an ER physician's NIHSS score agree with the neurologist's NIHSS score in the same patient at the same time in a real-world context? METHODS: Data was retrospectively collected from 1,946 patients being evaluated for AIS at Houston Methodist Hospital from 05/2016 - 04/2018. Triage NIHSS scores assessed by both the ER and neurology providers within one hour of each other under the same clinical context were evaluated for comparison. Ultimately, 129 patients were included in the analysis. All providers in this study were NIHSS rater-certified. RESULTS: The distribution of the NIHSS score differences (ER score - neurology score) had a mean of -0.46 and a standard deviation of 2.11. The score difference between provider teams ranged ±5 points. The intraclass correlation coefficient (ICC) for the NIHSS scores between the ER and neurology teams was 0.95 (95% CI, 0.93 - 0.97) with an F-test of 42.41 and a p-value of 4.43E-69. Overall reliability was excellent between the ER and neurology teams. CONCLUSION: We evaluated triage NIHSS scores performed by ER and neurology providers under matching time and treatment conditions and found excellent interrater reliability. The excellent score agreement has important implications for treatment decision-making during patient handoff and further in stroke modeling, prediction, and clinical trial registries where missing NIHSS scores may be equivalently substituted from either provider team.

BBox-Guided Segmentor: Leveraging expert knowledge for accurate stroke lesion segmentation using weakly supervised bounding box prior.

Stroke is one of the leading causes of death and disability in the world. Despite intensive research on automatic stroke lesion segmentation from non-invasive imaging modalities including diffusion-weighted imaging (DWI), challenges remain such as a lack of sufficient labeled data for training deep learning models and failure in detecting small lesions. In this paper, we propose BBox-Guided Segmentor, a method that significantly improves the accuracy of stroke lesion segmentation by leveraging expert knowledge. Specifically, our model uses a very coarse bounding box label provided by the expert and then performs accurate segmentation automatically. The small overhead of having the expert provide a rough bounding box leads to large performance improvement in segmentation, which is paramount to accurate stroke diagnosis. To train our model, we employ a weakly-supervised approach that uses a large number of weakly-labeled images with only bounding boxes and a small number of fully labeled images. The scarce fully labeled images are used to train a generator segmentation network, while adversarial training is used to leverage the large number of weakly-labeled images to provide additional learning signals. We evaluate our method extensively using a unique clinical dataset of 99 fully labeled cases (i.e., with full segmentation map labels) and 831 weakly labeled cases (i.e., with only bounding box labels), and the results demonstrate the superior performance of our approach over state-of-the-art stroke lesion segmentation models. We also achieve competitive performance as a SOTA fully supervised method using less than one-tenth of the complete labels. Our proposed approach has the potential to improve stroke diagnosis and treatment planning, which may lead to better patient outcomes.

Retrospective study of deep learning to reduce noise in non-contrast head CT images.

PURPOSE: Presented herein is a novel CT denoising method uses a skip residual encoder-decoder framework with group convolutions and a novel loss function to improve the subjective and objective image quality for improved disease detection in patients with acute ischemic stroke (AIS). MATERIALS AND METHODS: In this retrospective study, confirmed AIS patients with full-dose NCCT head scans were randomly selected from a stroke registry between 2016 and 2020. 325 patients (67 ± 15 years, 176 men) were included. 18 patients each with 4-7 NCCTs performed within 5-day timeframe (83 total scans) were used for model training; 307 patients each with 1-4 NCCTs performed within 5-day timeframe (380 total scans) were used for hold-out testing. In the training group, a mean CT was created from the patient's co-registered scans for each input CT to train a rotation-reflection equivariant U-Net with skip and residual connections, as well as a group convolutional neural network (SRED-GCNN) using a custom loss function to remove image noise. Denoising performance was compared to the standard Block-matching and 3D filtering (BM3D) method and RED-CNN quantitatively and visually. Signal-to-noise ratio (SNR) and contrast-to-noise (CNR) were measured in manually drawn regions-of-interest in grey matter (GM), white matter (WM) and deep grey matter (DG). Visual comparison and impact on spatial resolution were assessed through phantom images. RESULTS: SRED-GCNN reduced the original CT image noise significantly better than BM3D, with SNR improvements in GM, WM, and DG by 2.47x, 2.83x, and 2.64x respectively and CNR improvements in DG/WM and GM/WM by 2.30x and 2.16x respectively. Compared to the proposed SRED-GCNN, RED-CNN reduces noise effectively though the results are visibly blurred. Scans denoised by the SRED-GCNN are shown to be visually clearer with preserved anatomy. CONCLUSION: The proposed SRED-GCNN model significantly reduces image noise and improves signal-to-noise and contrast-to-noise ratios in 380 unseen head NCCT cases.