MixU-Net: Hybrid CNN-MLP Networks for Urinary Collecting System Segmentation

Segmenting the urinary collecting system based on preoperative contrast-enhanced computed tomography urography volumes is necessary for assisting flexible ureterorenoscopy. The urinary collecting system consists of complex elongated tubular structures and irregular treelike structures, making it challenging for precise segmentations using current deep-learning-based methods. Existing deep learning-driven methods face challenges in accurately segmenting the urinary collecting system from contrast-enhanced computed tomography urography volumes. In this work, we propose a novel MixU-Net by embedding global feature mix blocks. Particularly, the global feature mix blocks allow wider receptive fields based on fused multi-layer-perception and 3D convolutions across different dimensions. The experimental validations on the clinical computed tomography urography volumes demonstrate that our method achieves state-of-the-art in terms of dice similarity coefficients, intersection over union, and Hausdorff distance when compared with other methods that use pure convolutional neural networks or hybrid convolutional neural networks and Transformers. In addition, preliminary experiments conducted on the navigation system demonstrate the improved accuracy of the virtual depth maps when adopting the segmented urinary collecting system obtained by our MixU-Net.