Brain tumor sectionalization through semantic segmentation approach

Abstract

Accurate brain tumor detection and segmentation from magnetic resonance imaging (MRI) scans are vital for effective diagnosis, treatment planning, and patient monitoring. However, manual segmentation is time-consuming and subject to variability. As a result, it is a necessity for the development of automated solutions. Traditional convolutional neural network (CNN) approaches, such as U-Net and its variants, often face limitations in handling the high-resolution, complex patterns of MRI data. Traditional convolutional models struggle to generalize across diverse tumor characteristics and often fail to capture long-range dependencies, which are crucial for accurate segmentation and their computational inefficiency limits realtime applications. In the case of traditional transformer models, it often relies on fixed positional encodings to understand the spatial relationships between parts of an image.[25] There is also a need for adaptable models that do not rely on fixed positional encoding, enabling them to accurately segment tumors irrespective of their location within the brain. In this research we aim to address these challenges by implementing SegFormer, a Transformer-based architecture, and EfficientNetB4, a convolutional model, to enhance segmentation and classification accuracy for brain tumors.Here, we strategically adapted and integrated pretrained models, specifically SegFormer and EfficientNetB4 in order to create a robust framework for brain tumor segmentation and classification. Like other existing studies we did not focus on a single model. Rather, our approach combines SegFormer’s capability for precise pixel-level segmentation with EfficientNetB4’s efficient classification to improve overall diagnostic accuracy. In order to handle the specific challenges of high-resolution MRI data, we carefully tuned SegFormer, maintaining fine details and adapting to varying tumor characteristics, while using EfficientNetB4 to accurately distinguish between tumor and non-tumor images. Additionally, our method focuses on computational efficiency and real-time applicability by optimizing the models to ensure fast processing speeds, which is crucial for clinical application. In our experiment we demonstrated that SegFormer achieves superior segmentation performance, with a Dice score of 0.7961 and a Mean Intersection over Union (IoU) of 0.7382 which significantly outperforms other models like LinkNet, U-Net, and U-Net++, which recorded Dice scores of 0.3445, 0.2985, and 0.1575 respectively. Similarly, EfficientNetB4 achieved exceptional classification accuracy, with precision, recall, and F1-scores of 99% for both tumor and non-tumor classes, highlighting its reliability in distinguishing between the two. These results tell us that SegFormer’s efficient hierarchical encoder and pixel-level precision, combined with EfficientNetB4’s robust classification capabilities, offer a powerful and comprehensive solution for brain tumor segmentation and Lightweight MLP Decoder is computationally efficient for real time application.