MULTI HEAD ATTENTION-BASED LSTM AND GRADIENT-WEIGHTED CLASS ACTIVATION MAPPING FOR BRAIN TUMOR DETECTION USING MRI IMAGES

The presented of attention-based architectures in medical imaging has ushered in a novel era of precision diagnostics, mainly for the identification and classification of brain tumors. This research developing a novel knowledge distillation method that employs a tripartite attention mechanism within transformer encoder models to identify various brain tumor types utilizing magnetic resonance imaging (MRI). This study offerings a unique method for brain tumor identification, integrating Multi-Head Attention-based Long Short-Term Memory (MHA-LSTM) networks with Gradient-weighted Class Activation Mapping (Grad-CAM). The MHA-LSTM design utilizes multi-head attention to capture complex spatial-temporal relationships across consecutive MRI slices, enabling the model to focus on the most critical features. Grad-CAM is incorporated to provide visual explanations by highlighting key regions contributing to the model's predictions, improving both interpretability and clinical relevance. Experimental results demonstrate that the suggested technique surpasses conventional LSTM models in terms of accuracy, sensitivity, and specificity. Moreover, the Grad-CAM visualizations offer valuable insights into the model's decision-making process, fostering better understanding and facilitating future clinical validation. This method presented a robust and interpretable solution for brain tumor identification, advancing the application of deep learning in medical imaging.