Two-Stream nnU-Net: A Novel Architecture for Precise Tumor Segmentation in Medical Imaging
Chaymae El Mechal, Najiba El Amrani El Idrissi · 2023
This article introduces the Two-Stream nnU-Net, a novel architecture designed for automated tumor segmentation in medical imaging. Accurate tumor detection is vital for patient survival, but manual segmentation is challenging and time-consuming. This version of nnU-Net overcomes these challenges by adapting to various tasks, including preprocessing, network architecture, training, and post-processing, leveraging fixed parameters, interdependent rules, and empirical decisions to outperform traditional U-Net methods. By harnessing both structural and functional information from input images, nnU-Net independently trains two streams, combining their outputs for comprehensive segmentation. With the integration of skip connections and residual mappings, nnU-Net captures diverse features while reducing parameters, thereby enhancing convergence. A modified version of nnU-Net, featuring scalable components, demonstrates superior performance in tumor segmentation. This advancement marks a significant stride in the field of automated medical imaging, underscoring nnU-Net’s potential as a promising innovation.