Application of Artificial Intelligence in the Analysis of Magnetic Resonance Imaging for Identifying the Etiology in Stroke Patients
Ignacio Bozas1, Ignacio Germán Alfaro2, Ingrid Vanessa Daza Perilla3, Anabella Frances1, Agustin Allende4, Juan Jose Martin1
1Sanatorio Allende, 2IATE/OAC/UNC/CONICET, 3IATE/CRESST/UMBC/NASA-GSFC, 4Independent programmer
Objective:
Develop a machine learning algorithm for the etiological classification of ischemic strokes through the analysis of brain magnetic resonance imaging
Background:
Cerebrovascular disease is the leading cause of disability and the fourth leading cause of death worldwide, with 80% of cases being ischemic. Its multifactorial etiology is classified using systems such as TOAST. This study proposes an artificial intelligence model to analyze magnetic resonance images to early identify the underlying mechanisms of ischemic strokes, facilitating personalized treatment and optimizing resources in healthcare.
Design/Methods:
An algorithm was developed in Python to load and process brain MRIs of stroke patients, using the libraries NumPy, Pandas, Nibabel, and SciPy. The images, classified according to TOAST etiology, were extracted from patients at Sanatorio Allende and the Stroke Outcome Optimization Project. The images were normalized, generating data cubes with DWI, ADC, and FLAIR sequences. The data was split into 80% for training and 20% for evaluation in a convolutional neural network, and predictions and performance metrics were obtained.
Results:
A total of 1,180 neuroimages were analyzed. The DWI, ADC, and FLAIR sequences were selected, and the images were normalized to 130x130x150 pixels resolution. They were randomly divided into 944 images for training (80%) and 236 for testing (20%). The model, trained with supervised machine learning, achieved an average accuracy of 0.36, with a purity of 1.00 for category 3 and recalls of 0.55 and 0.42 for categories 1 and 5, respectively.
Conclusions:
Although the initial performance was low, it was observed that incorporating more examples improved accuracy. Technological limitations and the imbalance in the original sample affected the results. Nevertheless, it is expected that the inclusion of more images, adding metadata, and overcoming technological barriers will enhance the model's performance, moving closer to the effective application of this type of artificial intelligence in our clinical field.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.