CoalUMLP: Slice and Dice! A Fast, MLP-Like 3D Medical Image Segmentation Network

3D medical image segmentation tasks play a crucial role in clinical diagnosis. However, Handling vast data and intricate structures in Point-of-Care (POC) devices is challenging. While current methods use CNNs and Transformer models, their high computational demands and limited real-time capabilities limit their POC application. Recent studies have explored the application of Multilayer Perceptrons (MLP) to medical image segmentation tasks. However, these studies overlook the significance of local and global image features and multi-scale contextual information. To overcome these limitations, we propose CoalUMLP, an efficient vision MLP architecture designed specifically for 3D medical image segmentation tasks. CoalUMLP combines the strengths of CNN, Transformer, and MLP, incorporating three key components: the Multi-Scale Axial Permute Encoder (MSAP), Masked Axial Permute Decoder (MAP), and Semantic Bridging Connection (SBC). We reframe the medical image segmentation problem as a sequence-to-sequence prediction problem and evaluate the performance of our approach on the Medical Segmentation Decathlon (MSD) dataset. CoalUMLP showcases a state-of-the-art performance by significantly reducing the parameter count by 32.8% and computational complexity by 48.5%, all while maintaining a compact structure. Our results highlight the potential of CoalUMLP as a promising backbone for real-time medical image applications. It achieves a superior trade-off between accuracy and efficiency compared to previous Transformer and CNN-based models.