Abstract

To investigate the potential of radiomics features extracted from contrast-free brain Computed Tomography (CT) images to distinguish acute and chronic ischemic stroke lesions and to compare the performance of different machine learning algorithms. Data from 150 patients diagnosed with acute (n=75) or chronic (n=75) cerebral ischemia who underwent contrast-free brain CT between 2020 and 2025 were retrospectively analyzed. A total of 107 radiomics features, including shape, histogram, and texture matrix (GLCM, GLRLM, GLSZM, etc.), were extracted from the CT slices of all cases. Using these features, classification models were developed using the Support Vector Machine (SVM), logistic regression, k-nearest neighbor, Naive Bayes, decision tree, Random Forest, and Artificial Neural Network algorithms. Model performance was evaluated using 10-fold cross-validation with metrics such as Area Under the Curve (AUC), sensitivity, and specificity. The sex distribution was similar between the acute and chronic groups (male ratio 60% vs. 57.3%; p=0.781), but the chronic group had a higher mean age (74±14 vs. 63±12 years; p=0.0075). In the analysis of all 107 features, the Artificial Neural Network model achieved the highest AUC value (AUC=0.994). The highest sensitivity (99.6%) and specificity (99.3%) were obtained using the Random Forest model. In the classification using the five most important features determined by Information Gain—Skewness, High Gray Level Run Emphasis, Zone Entropy, Surface Area, and Gray Level Non-Uniformity Normalized—the Random Forest model also showed the best performance (AUC=0.988; sensitivity 99.6%; specificity 99.3%). Contrast-free CT radiomics features can accurately distinguish acute and chronic ischemic brain lesions. Radiomics-based artificial intelligence models may serve as powerful and promising tools for clinical decision support in determining the stroke stage, especially in cases where Magnetic Resonance Imaging (MRI) is not feasible. These findings highlight the potential of CT-based radiomics as a practical and rapid tool for stroke stage differentiation, especially in emergency settings where MRI is unavailable or contraindicated.

Key words: Ischemic stroke, radiomics features, computed tomography, machine learning, random forest, artificial neural network