U-Net++ architecture: A nested and dense design for multi-scale liver image segmentation from 2D CT scans
Sireesha Vadlamudi, Vimal Kumar, Debjani Ghosh · 2025
The mission of segmenting medical images is to split up an image into various areas based on some predefined criteria, which is essential. Convolutional neural networks have always been important and effective for segmentation tasks, especially for medical images, and the U-Net architecture is one of the most popular alternatives. In U-Net++, each pathway has a variety of depths and resolutions, allowing it to capture multiscale contextual information from the final image. The pathways are connected in a nested fashion, each receiving information from the first one, permitting us to more concluded segmentation of objects at different scales. To enhance the efficiency of U-Net++ for real-world and useful applications, we also introduce deep supervision during training. This allows for the lowering of redundant pathways at inference time without any problem in accuracy, reducing the computational complexity of the model. For validating our developed model we used the 3DIRCADB dataset for the liver segmentation from 2D abdominal Computed tomography scans and achieved good results. In gist, U-Net++ is an amazing power source architecture for image segmentation tasks, giving and suggesting a dense design that can capture multi-scale contextual information and preserve spatial information. It tells residual and skip connections to lower the vanishing gradient problem and deep supervision to improve efficiency.