Comparative Analysis of Machine Learning for Stroke Classification Using YOLOv11 Detection and a Radiomics-Based Two-Stage Model

Stroke is a leading cause of disability and death worldwide, including in Indonesia. Rapid and accurate diagnosis is crucial, especially during the golden period (3–4.5 hours). CT scans are the primary imaging modality, but manual interpretation is often limited by time, subjectivity, and radiologist availability. This study proposes a two-stage model integrating YOLOv11 for lesion detection and machine learning for classification, using radiomics for feature extraction. In the first stage, YOLOv11 detects lesions and generates bounding boxes, which serve as Regions of Interest (ROIs). In the second stage, radiomics features are extracted and classified using Naïve Bayes, Support Vector Machine (SVM), and Random Forest. Results show YOLOv11 achieved an overall mAP@50 of 0.732, with the highest performance in hemorrhagic stroke (0.741). Radiomics-based classification further improves stability, achieving accuracies of 0.97–0.99 and precision, recall, and F1 scores≥0.94. Among classifiers, SVM performed best, with a test accuracy of 0.97, a false positive rate of 1.23%, total error 0.0218, generalization gap -0.0117, variance 0.0002, standard deviation 0.003635, confidence interval 0.9708 (+/-0.0073), and consistent fold accuracy between 96.5–97.5%, indicating stability without overfitting. These findings confirm that the combination of the YOLOv11 two-stage model, radiomics, and SVM provides a robust approach to support stroke diagnosis.