Integrated U-Net 3+ with Decomposed Multi head Self-Attention Block and Filter Response Normalization for Brain Lesion Localization and Tracking using 3D MRI

In modern era, a fully automated system that uses 3D Magnetic Resonance Imaging (MRI) scans to segment brain tumors is very important for accurate diagnosis and better treatment planning. However, existing model face some challenges due to its morphology, variability and high computational cost. To overcome this problem, this paper proposed U-Net 3+ with Decomposed and Multi head Self-Attention Block with Filter Response Normalization (UNet3+DMSABFRN). The Decom-UNet3+ integrates decomposed convolution with a spatial attention mechanism and employs asymmetric convolutions and depth-wise separable convolutions to enhance directional awareness and improve inter channel feature extraction. In the MSAB layer, several self-attention heads are combined to capture complex relationships, and convolutions applied to adjust size of the output feature maps. After that, a Filter Response Normalization (FRN) layer is used for reducing the effect of batch size on the network. All MRI sequences are experimented with BraTS2020 dataset and by using Adam optimizer with a learning rate of 0.001 on T1 MRI sequence, the model achieved 99.90% accuracy and a Dice Similarity Coefficient (DSC) of 98%, showing that UNet3+DMSABFRN performs better than existing model like U-Net.