STROKECT-BENCH: Evaluating Convolutional and Transformer-Based Deep Models for Automated Stroke Diagnosis Using Brain CT Imaging

Abstract

Stroke detection from computed tomography (CT) images is an important research direction in computer vision. However, prior studies often use different preprocessing steps, model configurations, and evaluation protocols, making it difficult to compare results or assess architectural reliability. This paper presents an exploratory benchmark that evaluates representative convolutional neural networks (CNNs) and vision transformer (ViT) models under a unified experimental setting for binary stroke classification. STROKECT-BENCH is introduced as a standardized framework in which five CNNs and four transformerbased models are trained on the Brain Stroke CT Image dataset (1,551 normal and 950 stroke images) using identical preprocessing, augmentation, optimization parameters, and performance metrics. The results show that transformer models, particularly PVT-Small and Swin Transformer, achieve the highest accuracy and AUC, while EfficientNetB0 provides a strong balance between accuracy and computational efficiency. As an exploratory study, the findings aim to establish reliable baselines rather than clinical validation. STROKECT-BENCH offers a consistent evaluation reference for future work involving patient-level datasets, external validation, and multimodal stroke-analysis approaches.