The Multimodal MRI Brain Tumor Segmentation Using Modified Connected U-Net with a Guided Decoder

Brain tumor segmentation from multimodal imaging has remained a challenging task due to several image artifacts, the variable appearance of brain tumors, and class imbalance. The most advanced strategies for better segmentation performance are those based on deep-learning. This study presents a network architecture called modified Connected U-Net with Guided Decoder (GD), MCU-Net-GD, which combines two U-Net architectures and a guided decoder to enhance the segmentation performance further. Pre-trained ReNet-50 serves as the encoder block in first U-Net, and to efficiently capture semantic information, a second U-Net is used. The network uses ASPP block to gather contextual information. A guided decoder mechanism supervises the network's learning mechanism and creates comprehensive feature maps. The BraTS2019 dataset is used to examine the proposed network using DSC and IoU. Specifically, it obtained a dice score of 0.95 (WT), 0.95 (TC), and 0.91(ET) and IoU of 0.91 (WT), 90 (TC), and 0.83 (ET). The results demonstrate that the proposed network architecture performs better than the existing methods. This technique effectively reduces errors caused by weak contextual relationships by enhancing the integration of contextual information.