Discrete Wavelet-Based PSP-Integrated DenseU-Net Architecture for Brain Tumor Segmentation
摘要
Brain tumors pose significant health risks, and accurate segmentation of their subregions is crucial for effective diagnosis and treatment planning. Leveraging recent advancements in artificial intelligence (AI) and deep learning (DL), this paper proposes a novel architecture using discrete wavelet transform (DWT), modified pyramid scene parsing (PSP/PSP-DWT), dense convolution blocks that can handle the inherent challenges including presence of artifacts such as noise, bias fields, poor contrast, high intra-class and low inter-class variability, and guided decoder to deal with class imbalance in brain tumor segmentation. The proposed WPD-UNet model utilizes multi-scale feature maps generated using DWT, integrates spatial and frequency domain features, and employs DWT-pooling to enhance contextual relevance and feature saliency across multiple scales. This facilitates capturing both fine and coarse-level details. In addition the model incorporates the PSP mechanism with DWT-enhanced skip connections between encoder and decoder paths. A guided decoder using a weighted loss function addresses class imbalance, which is common in medical imaging. Comparative experiments using a mixture of LGG and HGG images on three publicly available datasets, BraTS2018, BraTS2019, and BraTS2020, show promising results, achieving competitive Dice scores compared to related approaches. The proposed model attains Dice scores of