Deep Learning-Based Brain Tumor Segmentation in MRI Images: A MobileNetV2-DeepLabV3+ Approach

Abstract

Introduction: Brain tumors (BTs) pose significant challenges in medical diagnosis and treatment due to their heterogeneity and complex characteristics. Accurate and precise segmentation of BTs in magnetic resonance images (MRIs) is crucial for effective treatment planning and patient care. In this study, we propose an ensemble deep-learning (DL) model to address the challenging task of BT segmentation. We aim to achieve accurate localization and delineation of tumor regions across different axial views. Material and Methods: The dataset used in this study consists of 3064 T1-weighted contrast-enhanced MRI images obtained from patients diagnosed with glioma, meningioma, and pituitary tumors. Image preprocessing techniques, including normalization and intensity transformation, were applied to enhance the contrast and standardize the intensity values. The DL model is based on the DeepLabV3+ architecture combined with three well-known deep convolutional neural networks as encoders: MobileNetV2, ResNet50, and XceptionNet. Results: The proposed ensemble model, with MobileNetV2 as the encoder, demonstrated superior performance in BT segmentation. The model achieved an average dice similarity coefficient of 0.938 and a global accuracy of 0.997. Compared to alternative models, MobileNetV2-DeepLabV3+ showed significant accuracy and segmentation precision improvements. Conclusion: The ensemble DL model, leveraging the strengths of MobileNetV2 and DeepLabV3+, offers a robust and efficient solution for accurate BT segmentation in MRI images. The model's ability to delineate tumor regions holds great promise for enhancing diagnosis and treatment planning in BT analysis. Future work will explore further fine-tuning techniques and evaluate the model's performance on larger datasets to assess its generalization capabilities.