Abstract

Brain haemorrhage remains a critical medical condition with high mortality and disability rates, necessitating timely and precise diagnosis. Traditional diagnostic approaches such as CT imaging often suffer from delays and subjectivity due to the reliance on radiologist interpretation. The main goal of this study is to create a deep learning-driven system that can automatically and reliably identify and categorize brain haemorrhages. The research addresses key challenges such as diagnostic delays, inconsistencies between medical evaluations, and the necessity for scalable and efficient diagnostic methods. This work aims to bridge the existing knowledge gap in real-time, generalized haemorrhage detection across diverse imaging scenarios. This work introduces a unique hybrid deep learning framework that integrates EfficientNetB0 with Bidirectional LSTM and Multi-Head Attention components. The EfficientNetB0 component efficiently extracts spatial features from CT images. These features are reshaped into temporal sequences and processed by BiLSTM to capture bidirectional dependencies. Subsequently, Multi-Head Attention is applied to focus dynamically on significant sequence segments, with residual connections enhancing stability. The training process employs the Adam optimization algorithm along with categorical cross-entropy loss enhanced by label smoothing for improved performance. Training is further regulated through dropout, early stopping, and learning rate scheduling—ensuring robustness. The combination of these elements enhances both the originality and performance of the suggested framework. Experimental results demonstrate that the model attains a classification accuracy of 98.03% and an F1-score of 0.99, surpassing traditional architectures like ResNet50, MobileNet, and DenseNet. Confusion matrix analysis demonstrates minimal false predictions, underscoring high sensitivity and specificity. These findings indicate that the model holds strong potential for use in clinical environments, especially where access to radiological expertise is limited. The integration of convolutional, sequential, and attention-based mechanisms significantly enhances diagnostic performance, offering an intelligent, a scalable approach aimed at enhancing diagnosis and treatment outcomes for individuals with potential brain haemorrhages.

Authors

Aruna Kokkula¹, P. Chandra Sekhar²
Maturi Venkata Subba Rao Engineering College, India¹, University College of Engineering, Osmania University, India²

Keywords

Brain Haemorrhage, Deep Learning, EfficientNetB0, Bidirectional LSTM, Multi-Head Attention, Medical Image Classification