Attention-Driven Sequential Feature Fusion Framework for Effective Brain Tumor Diagnosis

Abstract

Brain tumors are cells with abnormal growth patterns within the brain. They can appear in various forms and affect different areas of the brain. Brain tumors result from uncontrolled cell growth and are a leading cause of adult mortality. Early detection is crucial and significantly improves patient survival rates. MRI is effective for early detection, but traditional human inspection methods are inefficient for processing large datasets and time-consuming to handle large amounts of data. Our research proposes an advanced attention-based feature fusion method to address this issue. We utilized DenseNet201 and Xception as our base models due to their exceptional performance and proven efficacy in medical image classification tasks. These models excel at extracting robust and high-dimensional features, which are integrated into an advanced feature fusion framework to enhance diagnostic precision. This framework incorporates ConvLSTM layers and Convolutional Block Attention Modules (CBAM) to enhance feature attention and capture sequential dependencies. Our fusion technique leverages the strengths of these base models, further enhanced by CBAM attention mechanism. This approach represents a substantial advancement in medical imaging and provides an effective tool for diagnosing complex conditions. Our proposed model shows superior performance when evaluated on a widely recognized brain tumor dataset. It succeeds with an inspiring accuracy of 98.83%. This high accuracy demonstrates the potential of our method to advance early brain tumor detection significantly. By reducing misclassification rates, our model provides a reliable and efficient diagnostic tool that can improve patient outcomes. Additionally, we examined the generalizability of our proposed model on an additional Kaggle dataset consisting of 4 classes, which produced highly promising and effective results. In addition, visualization techniques such as Grad-CAM, feature maps, ROC curve and confusion matrix analysis are employed to interpret the model’s decisions and validate its effectiveness.

Keywords:Brain tumor; Classification; Feature fusion; Attention mechanism; Deep learning