Accurate Identification and Classification of Acute Ischemic Stroke Lesions Using Machine Learning Models and Region-Based MRI Features

Abstract

Introduction: Acute Ischemic Stroke (AIS) remains one of the leading causes of mortality and long-term disability worldwide. Early and precise lesion identification is critical for optimizing therapeutic interventions and improving patient outcomes. While radiological assessments like MRI play a vital role, manual interpretation can be time-consuming and prone to inter-observer variability. Consequently, there is a growing need for automated, machine learning-based approaches that enhance diagnostic efficiency and accuracy.

Objectives: This study aims to develop a robust and automated classification framework capable of accurately identifying and classifying AIS lesions using T1-weighted MRI scans. The objective includes evaluating the efficacy of different machine learning classifiers following a carefully designed feature extraction and pre-processing pipeline to support clinical decision-making.

Methods: The proposed framework leverages the ATLAS Release 1.1 dataset comprising annotated T1-weighted MRI scans of stroke patients. Pre-processing steps included skull stripping, spatial normalization, and brain parcellation using the Anatomical Automatic Labelling (AAL) atlas. Linear Discriminant Analysis (LDA) was applied to extract discriminative regional features. To address the class imbalance, the Synthetic Minority Over-Sampling Technique (SMOTE) was used. Multiple classifiers LinearSVC, Polynomial SVM, RBF SVM, Neural Network, AdaBoost, and QDA were trained and evaluated using stratified 10-fold cross-validation.

Results: Among the evaluated models, the Linear Support Vector Classifier (LinearSVC) demonstrated the highest classification performance, achieving an accuracy of 94.6%, along with superior precision, recall, and F1-score metrics. This highlights the effectiveness of LDA-driven feature extraction combined with class rebalancing techniques. Other classifiers showed competitive but comparatively lower performance, with the RBF SVM and Neural Network achieving accuracies of 91.2% and 90.5% respectively.

Conclusions: The integration of systematic pre-processing, atlas-based feature engineering, and classical machine learning offers a highly accurate and interpretable solution for AIS lesion classification. The results support the potential for deploying such a pipeline in clinical workflows to assist radiologists with rapid and reliable stroke assessment. Future work will focus on expanding the dataset, integrating multimodal inputs, and exploring deep learning-based feature extractors for improved generalizability.