Deep Learning-Based Brain Stroke Detection from CT Scan Images

This paper presented a hybrid EfficientNet-B0 + BiLSTM deep learning framework for automated three-class brain stroke classification from CT scans. The proposed architecture systematically combines the spatial feature extraction strength of a pretrained EfficientNet-B0 backbone — producing compact 256-D representations via 1×1 convolutional bottleneck compression — with the sequential spatial-temporal modelling capability of a Bidirectional LSTM that captures cross-regional contextual dependencies across CT feature map positions. The model achieves 97.14% test accuracy, 96.94% macro precision, 95.47% macro recall, and 96.18% macro F1-score on a 6,650-image brain CT dataset — outperforming Simple CNN (86.90%) by 10.24 percentage points and ResNet-50 (92.57%) by 4.57 percentage points, while achieving a 60% reduction in feature dimensionality. A class-weighted cross-entropy loss effectively mitigates the natural class imbalance between the Normal majority and the clinically critical Bleeding and Ischemia minority classes. Training dynamics confirm rapid convergence, strong generalisation, and absence of overfitting across 15 training epochs. The full-stack web application deployment (ReactJS + Flask + PyTorch) with real-time CT image upload and automated diagnostic report generation demonstrates the framework's readiness for integration into emergency radiology clinical workflows, offering immediate clinical decision support value.

 [1]      World Health Organization, "Stroke: Key Facts," WHO Technical Report, 2023. [Online]. Available: https://www.who.int/news-room/fact-sheets/detail/the-top-10-causes-of-death [2]     A. Smith, L. Chen, and R. Patel, "Deep learning-based ischaemic stroke detection from CT images," IEEE Transactions on Medical Imaging, vol. 43, no. 2, pp. 512–524, 2024. [3]     Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. 521, pp. 436–444, 2015. [4]     S. J. Pan and Q. Yang, "A survey on transfer learning," IEEE Transactions on Knowledge and Data Engineering, vol. 22, no. 10, pp. 1345–1359, 2010. [5]     J. Lim, K. Park, and S. Lee, "AI-powered lung cancer detection: VGG16 & CNN for CT image classification," Informatics (MDPI), vol. 12, no. 1, p. 15, 2025. [6]      M. Tan and Q. V. Le, "EfficientNet: Rethinking model scaling for convolutional neural networks," in Proc. 36th International Conference on Machine Learning (ICML), pp. 6105–6114, 2019. [7]     L. Wang, H. Zhang, and J. Liu, "Hybrid CNN-BiLSTM model for volumetric brain CT classification," IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 5, pp. 2189–2201, 2023. [8]      C. Rajeswari, B. Sathiyabama, and P. Devaki, "Innovations in stroke identification: ML-based diagnostic model using ResNet50 and classical classifiers," Computers in Biology and Medicine, vol. 175, p. 108441, 2024. [9]      X. Liu, T. Zhang, and W. Chen, "Deep residual global contextual feature fusion network for malignant nodule detection," Knowledge-Based Systems, vol. 289, p. 111512, 2025. [10]      R. Kumar and S. Verma, "Genetic algorithm-based feature optimisation for medical image classification," IEEE Access, vol. 11,pp. 34512–34527, 2023. [11]      S. Iqbal, H. Qureshi, and M. Ali, "Explainable AI for real-time medical image analysis and decision support," IEEE Access, vol. 11, pp. 78234–78249, 2023. [12]      K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016.