AI-driven smart e-healthcare for intracranial hemorrhage diagnosis using deep transfer learning and hybrid CNN-Bi-LSTM radiomics from CT scans

Brain hemorrhage, caused by ruptured arteries due to clots or high blood pressure, poses a life-threatening challenge that requires rapid and accurate diagnosis. Manual identification by radiologists is time-intensive and error-prone, particularly in the early stages. This article addresses the issue of brain hemorrhage identification, which is seen to be a difficult assignment for radiologists, especially during the initial stages of the hemorrhage. This study presents a Hybrid Deep Learning (HDL) framework combining Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), and transfer learning models (VGG16, ResNet50, DenseNet121) for classifying brain hemorrhage using CT scans. The framework is designed to perform effectively even with limited datasets. The effectiveness of the proposed approach is evaluated based on various performance evaluation parameters. The accuracy, recall, F1-score, precision, and ROC-AUC have been calculated with the observed maximum attainable accuracy of 93.36, while ROC-AUC is 98.34 under the DenseNet121. The hybrid imaging signature of CNN-BiLSTM achieved the highest performance of 89.67 for accuracy, with ROC-AUC 95.64. The neural network activation for different layers and the Gradient-weighted Class Activation Mapping (Grad-CAM) heatmap visualization of the best model have been observed. The proposed HDL models demonstrate high diagnostic accuracy and efficiency, highlighting their potential to assist radiologists in early hemorrhage detection, especially in time-critical scenarios.