<![CDATA[Stroke is still one of the world's leading causes of death and permanent disability, necessitating a quick and precise diagnosis in order to choose the best course of treatment. The purpose of this study is to examine how different batch size configurations affect the MobileNetV2 architecture's ability to classify stroke types from CT-scan brain pictures. The dataset comprises three categories Normal, Ischemic, and Bleeding sourced from Kaggle and RSUD Haji, East Java Province. The strategy to transfer learning was used utilizing pretrained ImageNet weights, with the network fine-tuned for stroke classification tasks. Experimental testing was conducted using three batch size configurations: 16, 32, and 64, while maintaining consistent hyperparameters for other training components. Among the assessment measures were accuracy, macro F1-score, and AUC (macro) to measure performance comprehensively. The results revealed that a batch size of 16 achieved the highest overall performance, with an accuracy of 96.14%, a macro F1-score of 96.15%, and an AUC of 99.62%, outperforming larger batch configurations. These findings indicate that smaller batch sizes enhance model generalisation and improve gradient update dynamics, enabling the CNN to better capture subtle patterns within CT-scan images. Thus, our study finds that the best trade-off between convergence speed and batch size is 16., model generalisation, and diagnostic accuracy, demonstrating the effectiveness of the MobileNetV2 architecture for automated stroke detection based on CT-scan imaging]]>
