Enhancing medical image segmentation through stacked u-net architectures with interconnected convolution layers

The effective integration of convolutional neural networks (CNNs), particularly as exemplified by the U-Net architecture, has led to notable advancements in medical image segmentation. Among various methodologies, the U-Net has demonstrated unparalleled proficiency across intricate medical image segmentation tasks, particularly when training and testing data emerge homogeneously from the same source domain. In this vein, our research introduces a pioneering paradigm by devising a novel stacked U-Net architecture. Our innovative model extends the conventional U-Net design through the strategic concatenation of two U-Net instances. Even though putting together multiple models is powerful, it comes with the problem of model parameters getting bigger, which makes them more likely to overfit or underfit because information can not get through the layers as easily during backpropagation. To counteract this quandary, we harness a mechanism of direct interconnections between the convolutional layer. To ascertain the robustness and supremacy of our proposed model, a trifecta of distinct medical image datasets is enlisted for comprehensive evaluation. Notably, a comparison of the conventional U-Net, Res-U-Net, and Inception-U-Net with cutting-edge U-Net architectures is made. The comparison clearly shows that our proposed architecture achieves superior segmentation performance across all datasets. In summation, our work augments the scientific discourse by presenting a potent avenue for advancing medical image segmentation paradigms.