Brain stroke detection using deep learning faces challenges due to imaging variations and subtle differences in CT scan features. This paper proposes a hybrid CNN model integrating EfficientNetB5 and ResNet101 in a parallel architecture to enhance classification accuracy. The model utilizes mixed precision training, AdamW optimization, and extensive data augmentation to improve generalization and robustness. Extracted features from both networks are merged and processed through dense layers with dropout and L2 regularization to prevent overfitting. The hybrid design leverages the strengths of both architectures, capturing fine-grained and deep semantic features effectively. Experimental results show superior performance over baseline models, demonstrating the potential of hybrid deep learning for reliable and efficient medical image classification.