RDCU-Net: A Multi-Scale Residual Dilated Convolution U-Net with Spatial Pyramid Pooling for Brain Tumor Segmentation

Tumors refer to abnormal growth of cells in the body. Early diagnosis of tumors plays a crucial role in improving treatment conditions , quality of life and patient survival. Deep learning methods are effective for medical image segmentation, but they struggle with tumors in magnetic resonance images (MRI) due to variations in intensity and appearance. Existing models like U-Net face challenges due to the integration of high-level and low-level features, leading to confusion. Our proposed model addresses the above issues by utilizing two techniques and fewer parameters compared to the existing methods, achieving higher accuracy. In the first technique, dilated convolution (DC) blocks with proportional rates are used to integrate high-level and low-level features. The second technique involves selecting dilated spatial pyramid (DSP) blocks, which increase the receptive field of features while maintaining their resolution, contributing to the network's generalization. The proposed model improves training, network depth, and feature extraction by incorporating a residual block. It outperforms the traditional U-Net model in terms of segmentation accuracy and network stability. We evaluated the model using the BraTS 2018 dataset, obtaining Dice coefficients of 0.906, 0.817, and 0.839 for the whole tumor (WT), the enhancing tumor (ET), and the tumor core (TC), respectively.