CNN Models for Analysis and Detection Ischemic Stroke Based on MRI Image

Stroke is first in Indonesia and among the top two causes of mortality worldwide. Stroke consists of 2 types, where ischemic stroke is the most common with an incidence rate of $\mathbf{8 7 \%}$. The length of time required to diagnose an ischemic stroke is the biggest obstacle, especially in emergency conditions for treating this disease. This research presents novelty by combining classification and segmentation for analysis, while other studies usually only focus on one. This approach allows for more complete results by differentiating ischemic strokes from controls and determining their location. This research compares four CNN architectures, namely MobileNetV2 and VGG19 for classification, and U-Net and SegNet for segmentation. This research also compared several optimizers, namely SGD, and RMSprop, and variations in learning rates of 0.01 and 0.001. This research also uses transfer learning with $30 \%$ and $60 \%$ training rates. This research aims to find the most accurate model to support timely and appropriate decision-making regarding the detection of ischemic strokes. The research results show that the best classification model is MobileNetV2 with SGD optimizer, learning rate of 0.01, and trainable $30 \%$, better than VGG19 with an accuracy of $97 \%$. U-Net is the best segmentation model compared to SegNet, with the RMSprop optimizer, $\mathbf{9 8 \%}$ accuracy, and 0.001 learning rate achieved. This research reveals the accuracy of the best models in classifying and segmenting, which can recognize ischemic stroke and control image test data well.