<![CDATA[Stroke is a cerebrovascular disorder that can cause tissue damage, disability, and death. Early diagnosis is important in stroke management, particularly in the acute phase. Non-contrast CT scan (NCCT) is a widely available imaging modality, but its images often show subtle findings that are difficult to identify visually. This study aims to develop a stroke classification model on NCCT images using the Vision Transformer (ViT) method, analyze the effect of skull removal preprocessing on classification results, and evaluate model performance. The data were obtained from the Acute Ischemic Stroke Detection (AISD) dataset, consisting of non-contrast CT scan sclices images from normal (without lesion) and ischemic stroke (with lesion) classes. The research method included data collection, image preprocessing using two scenarios, namely without skull removal and with skull removal, CLAHE and resize implementation, data augmentation, dataset splitting, ViT model training, and evaluation. The results showed that non-pre-trained ViT-Large model on this dataset reached a best performance of 66.53%, with an average of around 65%. These results indicate that the model is capable of learning basic stroke patterns quite well but still has limitations in generalization due to the limited amount of training data. This finding is reinforced by the results of a comparative test using a transfer-learning-based ViT model with pre-trained ImageNet weights, which was able to increase accuracy to 85.96%. Thus, the Vision Transformer method has the potential for high accuracy in supporting stroke diagnosis, but its performance is highly dependent on the availability of large amounts of training data]]>
