Predicting Brain Tumor Region from MRI FLAIR Images using Ensemble Method

The conventional approach to tumor detection primarily hinges on the expert evaluation of physicians, a process susceptible to time constraints and inherent inaccuracies, thereby posing imminent threats to patients’ well-being. However, the integration of the Fluid Attenuated Inversion Recovery (FLAIR) modality offers a transformative avenue. It empowers medical practitioners to gain profound insights into the extent of tumor infiltration, representing a significant advancement in diagnostic capabilities and potential treatment strategies. This study delves into the realm of brain tumor assessment and prediction using various deep learning CNN architectures like ResNet50 U-Net, EfficientNetB3 U-Net, and InceptionV3 U-Net. The U-Net architecture, in particular, stands out for its adaptability, allowing the incorporation of several neural network pre-trained models within a single framework. The investigation of ensemble modeling resulted from a focus on improving the effectiveness of the segmentation approach. This method makes use of the previously described models’ combined advantages, enabled using an average of softmax. This cooperative endeavor resulted in the Ensemble model’s development, which surpasses its individual analogs in terms of performance measures. Combining multiple models with their unique strengths and weaknesses enhances overall ensemble robustness. Diverse segmentation models employing different features benefit from ensembling, capturing a wider range of patterns for improved performance. In this context, the ensemble model emerges as the most effective in terms of segmentation performance, achieving a Dice Similarity Coefficient of 93.29% and an Intersection of Union score of 87.65% on the datasets. These metrics quantify the overlap between actual and predicted regions, showcasing the superiority of the ensemble model which outperforms the state-of-the-art methods.