Hybrid DWT–U-Net Framework for Brain Tumor Segmentation on the BRATS-2020 Dataset

Abstract

Brain tumors are abnormal growths in brain tissue that significantly impact neurological functions. Accurate segmentation of tumor regions in magnetic resonance imaging (MRI) is crucial for diagnosis, treatment planning, and patient monitoring. However, manual delineation is time-consuming and subject to inter-observer variability.

This study presents a hybrid segmentation framework that integrates the Discrete Wavelet Transform (DWT) with a U-Net architecture for automated brain tumor segmentation. The wavelet transform is employed to extract multi-scale frequency features and enhance tumor boundaries while reducing noise. These refined features are subsequently fed into the U-Net, which learns hierarchical spatial representations to generate precise segmentation masks.

The proposed method was evaluated on the BRATS-2020 dataset, which includes multi-modal MRI sequences (FLAIR, T1, T1ce, and T2). Quantitative performance was assessed using the Dice coefficient, Jaccard index, Precision, Accuracy, Sensitivity, and F1-score. Experimental results demonstrate that the hybrid DWT–U-Net approach outperforms the baseline U-Net by improving boundary delineation, minimizing false positives, and enhancing segmentation accuracy.

Overall, this hybrid methodology offers an efficient and reproducible framework for automated brain tumor segmentation, combining the strengths of classical signal processing and deep learning to support clinical diagnosis and treatment planning.