Unified Deep Learning Method for Accurate Brain Tumor Segmentation Using Vertical Voxel Grouping and Wavelet Features

Abstract

Brain tumor segmentation plays a vital role in medical imaging, enabling accurate diagnosis and guiding treatment decisions. Despite notable progress driven by deep neural networks (DNNs) and multi-parametric magnetic resonance imaging (mpMRI), the complexity and heterogeneity of tumor tissues make precise segmentation a persistent challenge. In this paper, we propose a novel method that integrates Vertically grouped Voxel Feature Extraction (VFE), wavelet-based multi-resolution detail enhancement, and a modified UNet-VGG16+ architecture. The VFE component enhances tumor region contrast and suppresses irrelevant background areas by grouping column-wise voxel intensities within each slice. As a result, the average image contrast is increased by 23.78%, thereby improving the ability of Deep Neural Networks (DNNs) to focus on tumor regions. The wavelet-based enhancement captures multi-resolution details to more clearly delineate tumor boundaries while also reducing noise. The UNet-VGG16+ architecture leverages transfer learning to efficiently process these enhanced features for accurate segmentation. Extensive experiments on the BraTS21 dataset demonstrate that the proposed method achieves a mean Dice score of 94.69%, with segmentation accuracies of 93.3%, 93.1%, and 94.4% for Enhancing Tumor (ET), Whole Tumor (WT), and Tumor Core (TC), respectively. Comparative evaluations show consistent and statistically significant improvements over state-of-the-art models (p< 0.001). Further validation on the BraTS18 dataset confirms the model's generalizability. These results highlight the effectiveness of combining spatially structured voxel aggregation with frequency-domain analysis for robust and high-precision brain tumor segmentation.