Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-10950337-B1
Publication Date
2021-03-16
Expiration Date
Abstract
Disclosed are an ASPECT score estimating system and a method thereof which estimate an ASPECT score which is a factor for identifying the stroke diseases from the brain CT images and a configuration including a pre-processing step of normalizing and standardizing a feature of an image dataset; a segmenting step of separating each lesion from an CT image which is classified into a supra ganglionic level and a ganglionic level; and a determining step of determining whether the lesion is a stroke by independently building a neural network which learns a positive/negative image for each lesion is provided to estimate an ASPECT score which is an objective indicator for diagnosing a condition of the stroke patient using a brain CT image of the patient.
Core Innovation
The invention relates to an Alberta stroke program early computed tomography (ASPECT) score estimating method based on a brain computed tomography (CT) image of a stroke patient. The method estimates an ASPECT score by segmenting stroke lesions into a supra ganglionic level and a ganglionic level, and uses image-processing and neural-network based determination of whether each lesion is a stroke through positive/negative decisions.
The pre-processing step normalizes and standardizes an image dataset by removing noise included in the entire brain CT image using convolution with a Gaussian blur. A skull finding step finds a skull from the brain CT image, and an alignment step constantly aligns a position and a rotation degree by aligning an image position with respect to a center point of the found skull. A horizontal inverting step is performed in accordance with a lesion-side, and the CT image is subjected to pre-processing to be normalized and standardized such that a left brain is a lesion and a center vertical line is a symmetry line of the brain.
The invention further specifies skull-focused processing by auto thresholding to search an adaptive threshold value based on distribution of pixel values so that only pixel information corresponding to the skull remains, followed by contour finding to detect an edge based on an image after thresholding. Skull ellipse information acquiring then searches skull ellipses from the image having only edge information to obtain internal and external ellipse information, which is used to crop each lesion image from the entire image in accordance with segmentation results.
In the determining step, supervised learning is performed on the cropped image using a neural network configured for every lesion, where independently built neural networks for each lesion determine positive/negative and the final ASPECT score is calculated based on the number of neural networks that determine the lesion as positive, with independent neural networks for the supra ganglionic level and the ganglionic level.
Claims Coverage
The provided content includes one independent claim. The claim contains inventive features covering normalization and standardization pre-processing of brain CT images, lesion segmentation into supra ganglionic and ganglionic levels, lesion cropping and supervised deep learning for lesion-wise positive/negative decisions, and independent neural networks for lesion determination and final ASPECT score calculation.
Gaussian blur noise removal for CT normalization
Removing noise included in the image by convoluting the entire brain CT image of the patient with a Gaussian blur.
Skull finding and skull-centered alignment
Finding a skull from brain CT image and constantly aligning a position and a rotation degree of the dataset by aligning a position of the image and rotating the image with respect to a center point of the found skull.
Lesion-side horizontal inversion and symmetry normalization
A horizontal inverting step in accordance with a lesion-side, wherein the CT image is subjected to the pre-processing step to be normalized and standardized as an image centered such that a left brain is a lesion and a center vertical line of the image is a symmetry line of the brain.
Adaptive-threshold skull masking with contour and ellipse detection
Searching an adaptive threshold value in consideration of distribution of pixel values of the segmented image and inducing only pixel information corresponding to the skull to remain, followed by detecting an edge based on an image after thresholding and acquiring internal and external ellipse information of the skull by searching skull ellipses from the image only having edge information.
Lesion cropping from entire image based on segmentation results
Cropping each lesion image from the entire image in accordance with the segmentation result of the segmenting step.
Supervised learning on cropped lesion images
Performing supervised learning on the cropped image with a neural network configured for every lesion including positive/negative information.
Independent lesion-wise neural networks and ASPECT score calculation
Learning and classifying in a neural network independent for every lesion and calculating a final ASPECT score based on the number of neural networks which determine the lesion to be positive.
Separate neural networks for supra ganglionic level and ganglionic level
Building an independent neural network for the supra ganglionic level and the ganglionic level.
CNN configuration with six hidden layers, dropout, and ReLU
The neural networks are independently built based on a convolutional neural network (CNN), are configured by six hidden layers and two dropout layers for preventing overfitting, and use a rectified linear unit (ReLU) as an activation function.
The independent claim covers an ASPECT score estimating method that uses normalized and standardized brain CT preprocessing, skull-focused thresholding and edge/ellipse detection, lesion segmentation into supra ganglionic and ganglionic levels, cropped lesion learning with independent neural networks, and final ASPECT score calculation from positive lesion determinations.
Stated Advantages
Reduces interrater scoring variability.
Documented Applications
Estimating an Alberta stroke program early CT (ASPECT) score from brain CT images of a stroke patient.
Evaluation/verification of lesion-level and level-wise network performance using Bland-Altman plot and reported accuracies for M4 supra-ganglionic level and ganglionic vs supra-ganglionic levels.
Interested in licensing this patent?