PLKA-nnUNet: A Parallel Large Kernel Convolutional Attention nnUNet Framework for Rodent Skull Stripping

Skull stripping is a vital preprocessing step in magnetic resonance imaging (MRI), but it faces challenges such as low contrast between brain tissue and boundaries and uneven magnetic fields, particularly for rodents. Currently, this process heavily relies on manual expertise. In this study, we developed a nnUNet-based framework, PLKA-nnUNet, for automating skull stripping operations in mouse and rat MRI scans. The primary goal of this framework is to minimize preprocessing efforts and offer an adaptable, out-of-the-box solution for various MRI image resolutions. A deep learning neural network is employed to achieve fully automated skull stripping, resulting in an efficient and lightweight model. To validate our method, we trained and evaluated the network using our rodent brain dataset and additional multicenter data. Experimental results show that our method achieves the highest Dice coefficient (0.9901), Hausdorff distance (4.3851), and Jaccard index (0.9874), with low model parameters and computational complexity.