Automated Multimodal Brain Tumor Segmentation and Localization in MRI Images Using Hybrid Res2-UNeXt

The early diagnosis and precise localization of brain tumors are vital for improving and saving patients' lives. In the realm of medical image analysis, a pressing challenge is to develop a reliable method for effectively separating tumor regions from healthy tissues in brain MRI scans. This research introduces an automated brain tumor segmentation approach utilizing a hybrid model, combining a residual network and UNeXt (referred to as hRes2-UNeXt). In this architecture, the U-Next model functions as an encoder, while a residual network serves as an encoder to mitigate the zero-gradient problem. Additionally, skip connections are employed between residual and convolutional blocks to expedite the training process. The evaluation of this approach on the BRATS 2021 dataset yielded promising results, with mean dice scores of 0.9157 for the tumor core (TC), 0.9320 for the whole tumor (WT), and 0.9226 for the enhancing tumor (ET). Comparative analysis with state-of-the-art techniques demonstrates the substantial enhancement in brain tumor sub-region segmentation accuracy achieved by the hybrid Res2-UNext model.