Stroke Detection and Diagnosis - Research Program


Project: Stroke Lesion MRI Segmentation for Patient Outcome Prediction

Representative stroke infarct segmentation using diffusion MRI (top row), apparent diffusion coefficient map (middle row) as input to AI model and the segmented results (bottom row).

Infarct volume is an independent predictor of 90-day modified Rankin Score (mRS) outcome in acute ischemic stroke (AIS). Deep learning image segmentation of infarct volume has shown promising results recently. Unfortunately, these all are based on traditional convolutional neural networks and can generate unpredictable results when input images are mirrored, flipped, or rotated. Manual measurement is laborious, and its variance is hard to quantify. This research aims to develop automated segmentation methods with built-in robustness for image rotation and mirror flipping. We have validated the accuracy over a large range of lesion sizes in more than 1,000 stroke patients, including a cohort of 175 cases who underwent thrombectomy. The automated method can segment acute stroke volume, even for infarct in choroid plexus. Segmented acute stroke volume is a significant prognostic indicator of 90-days outcome measured by mRS even after adjusting the effects of patients age and baseline NIH stroke scale (NIHSS).

Project: Deep Denoising of non-contrast Brain CT Scans

(a) Original NCCT, (b) denoised CT, (c) bias field corrected denoised CT, (d) diffusion MRI. The red crosshair corresponds to acute ischemic stroke location that has been irreversibly damaged.

In acute ischemic stroke, tissue ischemia caused an increase of tissue water. During the first 3 hours of acute ischemic stroke, we typically see 2-4 Hounsfield Units decrease in the stroke region on non-contrast CT (NCCT) images. Compared to image noise normally around 4-7 Hounsfield Units at full-dose radiation, early acute stroke detection thus presents a challenge for disease diagnosis and interpretation. This research project investigates a novel CT enhancement and noise reduction method using advanced deep learning algorithms to boost the image signal-to-noise (SNR) level comparable to high-quality MRI for improved neurological disease detection. In addition, the proposed deep denoising method is based on image-space and can be applied to any non-contrast head CT scans from varying vendors to improve disease detection. It is especially useful in emergency settings when MRI is not available. A prospective clinical trial on the technology in acute ischemic patients is ongoing at Houston Methodist.

Project: Multimedia Intelligence for Stroke Screening and Assessment

Dataflow diagram of the video/audio stream network fusion process for stroke/TIA prediction using the feature output from the ResNet-34 before the final fully-connected layer.

Acute ischemic stroke is a common but challenging disease to diagnose in the emergency setting. Facial asymmetry is a frequent finding in patients with stroke. This work proposes an AI framework for screening and assessing facial asymmetry caused by stroke using video analysis on patient videos from an emergency room setting, performing spatiotemporal analysis on facial asymmetry and providing a clinical decision tool to support emergency diagnosis. Clinical datasets for this study were acquired by the Eddy Scurlock Comprehensive Stroke Center at Houston Methodist with susceptible stroke patients and healthy volunteers as controls. Subjects are asked to first say the phrase, “Today is a sunny day in Houston, Texas,” and describe a picture known as the “Cookie Theft” picture while being recorded. Videos are labeled with the clinical impressions on whether or not patients have a stroke or transient ischemic attack (TIA), as well as discharge diagnosis for validation. Acquisition of facial data in natural settings makes our work more robust and useful for clinical use.