Performance and Explainability of Modern Deep Learning Models for Brain Tumor MRI Segmentation

Brain tumor segmentation from magnetic resonance imaging (MRI) is essential for accurate diagnosis, treatment planning, and disease monitoring. While convolutional neural networks (CNNs) have achieved remarkable success in this task, recent Transformer and Mamba-based architectures have emerged as promising alternatives. However, their effectiveness and interpretability for brain tumor segmentation remain under-explored. This paper presents a comprehensive comparative study of modern deep learning models, including U-Net, Attention U-Net, DeepLabV3+, encoder-based U-Net variants, TransUNet, Simplified Mamba U-Net, and Mamba U-Net, using a publicly available brain MRI dataset. Performance is evaluated using Dice, IoU, Precision, Recall, F1-score, Pixel Accuracy, ROC-AUC, model complexity, inference latency, and model size. In addition, Explainable Artificial Intelligence (XAI) techniques are employed to visualize and compare model decision-making behavior. Experimental results show that lightweight CNN-based architectures, particularly EfficientNet-B0 U-Net, achieve the best balance between segmentation accuracy and computational efficiency, while Transformer and Mamba-based models offer competitive performance. The XAI analysis further highlights differences in feature localization across architectures, providing valuable insights into model reliability and supporting informed model selection for brain tumor MRI segmentation.