E-070 Machine vision and machine learning radiomic texture analysis for neurointervention

Abstract

Introduction Machine learning and machine vision analysis are set to revolutionize medical image analysis within radiology. Initial results have shown there is much promise from the ability of machine learning algorithms to access data in ways not available to the human eye. One of these areas is radiomic texture analysis, which generates features from data based on the pixel/voxel histogram intensity values. Previous work has shown promise in identification of stroke before it is clinically apparent. Methods 82 patients triaged for stroke with Multiphase CTA head and neck to assess for acute ischemic stroke, large vessel occlusion and candidacy for thrombectomy and were identified and and included in this retrospective cohort IRB approved study.Inclusion criteria included presence of an M1 occlusion without tandem occlusion, or no occlusion and non-contrast head CT images available in the picture archiving system. Noncontrast head CT images were evaluated and the basal ganglia were manually segmented with KNIME image processing software using the single slice with the largest volume of basal ganglia included, as selected by a Neurointerventional radiology fellow. A total of 102 image features were extracted, loosely grouped into haralick, tamura, and pixel intensity histogram based features for each the left and right basal ganglia. Subsequently the image features were subtracted left from right, and geometric center data was discarded. Boostrap sampling or sampling with repletion with 10 repetitions was used to stratify the data into groups for training and testing of the machine learning algorithms. Training data was further resampled using cross validation with 4 folds for feature selection using a forward selection algorithm, retaining the best 14 features. Naive bayes algorithm was used for prediction with laplace correction. Next only the thrombectomy cases were included and the results were used for prediction of initial dichotimized NIHSS scores. Finally the results were used to predict dichotimized discharge mRS scores. Results Results were as follows: Stroke present accuracy: 63.09%±4.03%, AUC: 0.648±0.074, sensitivity: 56.37%±16.23%, specificity: 66.74%±8.22%, Positive predictive value: 50.60%±5.98%, Negative predictive value: 72.02%±6.52% Initial NIHSS accuracy: 60.63%±12.67% AUC: 0.605±0.183 sensitivity: 75.05%±21.57% specificity: 39.53%±27.10% positive_predictive_value: 68.66%±15.36% negative_predictive_value: 45.83% Dicharge mRS AUC: 0.545±0.210 In comparison to a similar result by chance this result has a p value of less than 0.001 based on the receiver operating characteristic area under the curve, or ROC AUC for categorization of stroke presence. Initial NIHSS prediction ROC AUC p value was 0.0038, whereas the p value for ROC AUC for dicharge mRS was p=0.25. Conclusion Although much work remains to be done, These promising results show feasibility of LVO classification and feasibility of clinical outcomes classification from sparse data. Work with a convolutional neural network approach is ongoing, but multi-institutional collaboration is needed. Disclosures S. Arndt: None. J. Milburn: None. G. Bennett: None.