System and methods of prediction of ischemic brain tissue fate from multi-phase CT-angiography in patients with acute ischemic stroke using machine learning
The invention relates to systems and methods for predicting ischemic brain tissue fate from multi-phase CT-angiography. More specifically, systems and methods are provided that enable meaningful prediction of core, penumbra and perfusion from mCTA images using software that has been trained via machine learning to interpret mCTA images.
1 . A method of quantifying core and/or penumbra from a plurality of current multi-phase computed tomography (mCTA) images of a patient comprising the steps of:
i. introducing the plurality of mCTA images into a prediction model, the prediction model derived from historical computed tomography perfusion (CTP) image data and CTP study data that quantified Time to Maximum (Tmax), cerebral blood volume (CBV) and cerebral blood flow (CBF) from the historical CTP image data and wherein the prediction model fits the current mCTA images into the prediction model to predict core and/or penumbra from the current mCTA images.
2 . The method as in claim 1 where the historical CTP image data further comprises patient treatment data, patient post-treatment follow-up images and patient outcome data and where the prediction model fits a current patient core/penumbra prediction to the patient outcome data to obtain a prediction of outcome of the current patient.
3 . The method as in claim 2 where the patient treatment data includes surgical procedure data whether undertaken or not.
4 . The method as in claim 1 further comprising the steps of calculating prediction maps and displaying the prediction maps on a display system and where the prediction maps include core and/or penumbra as core and/or penumbra prediction maps.
5 . The method as in claim 1 further comprising the steps of calculating an outcome score for a current patient based on a calculation of total core and/or penumbra and/or diffusion, fitting the total core and/or penumbra and/or diffusion to past patient data having outcome data and displaying the outcome score on a display system.
6 . The method as in claim 1 further comprising the step of predicting core/penumbra within 10 minutes of initially obtaining current mCTA images.
7 . A method of building and training a machine learning database to enable prediction of any one of or a combination of core, penumbra and perfusion status from multi-phase computed tomography (mCTA) images comprising the steps of:
i. introducing historical patient data into a database, the historical patient data including images from multiple computed tomography perfusion (CTP) studies and treatment follow-up images;
ii. analyzing the historical patient data to extract features of interest relating to occlusion location, core, penumbra and perfusion;
iii. introducing historical mCTA patient data in the database, the historical mCTA patient data including multiple sets of mCTA images and testing the sets of mCTA images obtained in step i using a machine-learning algorithm, where each set of mCTA images include phases of images and follow-up images;
iv. deriving a classifier prediction model from step iii; and,
v. introducing a single set of mCTA image data into the prediction model from step iv and analyzing the mCTA image data to produce any one of or a combination of a core, penumbra and status prediction probability map for the mCTA image data.
8 . The method as in claim 7 where the historical patient data includes data from patients having undergone reperfusion and patients not having undergone reperfusion.
9 . The method as in claim 8 where the historical mCTA patient data includes data from patients having undergone reperfusion and patients not having undergone reperfusion.
10 . The method as in claim 7 where the prediction model calculates predicted core volume.
11 . The method as in claim 7 where the prediction model calculates predicted penumbra volume.
12 . The method as in claim 7 where the prediction model calculates predicted tissue perfusion status.
13 . The method as in claim 7 where the prediction model determines follow-up infarct volume and utilizes the follow-up infarct volume as a reference standard for step v.
14 . The method as in claim 7 where the step of feature extraction includes the steps of analyzing density and acquisition time of features.
15 . The method as in claim 14 where density and acquisition time analysis includes for each voxel:
i. calculating average and standard deviation of Hounsfield Units (HU) across each phase of mCTA images;
ii. calculating a coefficient of variance of HUs for each phase of mCTA images;
iii. calculating slopes of HUs between any two phases of mCTA images;
iv. determining peak of HUs across the phases of mCTA images; and,
v. determining a time peak of HUs.
16 . The method as is in claim 15 where features are calculated in the neighborhood centered at each voxel at different scales.
17 . The method as in claim 15 further comprising the step of comparing the mCTA prediction probability map against follow-up images to test the accuracy of the model.
18 . A method of building and training a machine learning database and model to enable prediction of any one of or a combination of core, penumbra and perfusion status from sets of multi-phase computed tomography (mCTA) images and sets of computed tomography perfusion (CTP) images, where each set of mCTA images include phases of images and follow-up images, the method comprising the steps of:
i. introducing historical patient mCTA and CTP images into a database and analyzing the mCTA and CTP images to extract features of interest relating to occlusion location, core, penumbra and perfusion;
ii. testing multiple sets of mCTA images against patterns obtained in step i using a machine-learning algorithm;
iii. deriving a classifier prediction model from step ii;
iv. introducing a single set of mCTA image data into the prediction model from step iii and analyzing the mCTA image data to produce any one or more of a core prediction, a penumbra prediction, and a perfusion status prediction, the any one or more of the core prediction, the penumbra prediction, and the perfusion status prediction forming an MCTA prediction probability map for the mCTA image data; and,
v. comparing the mCTA prediction probability map against follow-up images to ascertain the accuracy of the model.
19 . The method as in claim 18 wherein steps ii and iii includes two-stage training including a first penumbra stage that derives a penumbra area and a second core stage that derives a core area.
20 . The method as in claim 18 wherein the machine learning model comprises one of a random forest, a support vector machine, a neural network, or a k nearest neighbor model.
21 . The method as in claim 18 wherein the features of interest relating to occlusion location, core, penumbra and perfusion, are identified by any one of or a combination of first-order statistics including mean and histogram of HU values, and texture features including gray-level co-occurrence matrix and gray level run length matrix.
22 . The method as in claim 18 wherein the features of interest relating to occlusion location, core, penumbra and perfusion are calculated at different scales for a given voxel corresponding to the axial imaging and where the features of interest are calculated at low, median, and high resolution scales.
23 . The method as in claim 18 wherein the features of interest contributing to occlusion location, core, penumbra and perfusion are automatically selected using a feature selection module utilizing any one of or a combination of univariate selection, feature importance, and correlation matrix with heatmap.
24 . The method as in claim 18 wherein at least one probability map is thresholded to generate infarct core and/or penumbra and/or perfusion volume for an axial imaging slice.
25 . The method as in claim 24 wherein morphological operations including dilation and/or erosion and component analysis are applied after thresholding to remove isolated islands.
26 . The method as in claim 18 wherein the model enables prediction of any one of or a combination of core and penumbra within a multiple label machine learning model including a core label, penumbra label and normal tissue label.
27 . The method as in claim 18 further comprising the step of inputting historical patient meta data including age, sex, NIHSS, ASPECTS, and occlusion site.
28 . A method of predicting a plurality of contrast enhanced volumes in a brain scan image comprising the steps of:
from a series of multi-phase computed tomography (mCTA) images from a stroke patient and a plurality of historical images from patients having undergone non-contrast computed tomography (NCCT) and computed tomography perfusion (CTP) study,
i. comparing signal intensity fluctuations of voxel data of the mCTA images against corresponding voxels from the historical CTP images and time synchronizing a plurality of mCTA volumes to a plurality of CTP volumes;
ii. from time synchronized mCTA and historical CTP volumes, comparing corresponding voxels from the mCTA images and historical CTP images and finding at least one match of historical CTP images; and,
iii. utilizing the at least one match of historical CTP images as basis for predicting a contrast enhanced volume for the mCTA images.
29 . The method as in claim 28 further comprising the step of building and displaying at least one predictive maps showing a combination of core and penumbra and/or perfusion.