This paper proposes UNet 3+, a novel full-scale connected UNet for medical image segmentation. UNet 3+ improves upon previous architectures by incorporating full-scale skip connections and deep supervision to better utilize multi-scale features. The full-scale skip connections combine low-level details with high-level semantics from feature maps across different scales, while deep supervision learns hierarchical representations from aggregated full-scale feature maps. Additionally, a hybrid loss function and classification-guided module are introduced to enhance organ boundary detection and reduce over-segmentation in non-organ images. The proposed method achieves improved segmentation accuracy and computational efficiency, with fewer parameters compared to existing methods. Experiments on liver and spleen datasets show that UNet 3+ outperforms previous state-of-the-art approaches, achieving higher Dice coefficients and producing more accurate, coherent segmentation results. The method is particularly effective for organs that appear at varying scales. The code is available at https://github.com/ZJUGiveLab/UNet-Version.This paper proposes UNet 3+, a novel full-scale connected UNet for medical image segmentation. UNet 3+ improves upon previous architectures by incorporating full-scale skip connections and deep supervision to better utilize multi-scale features. The full-scale skip connections combine low-level details with high-level semantics from feature maps across different scales, while deep supervision learns hierarchical representations from aggregated full-scale feature maps. Additionally, a hybrid loss function and classification-guided module are introduced to enhance organ boundary detection and reduce over-segmentation in non-organ images. The proposed method achieves improved segmentation accuracy and computational efficiency, with fewer parameters compared to existing methods. Experiments on liver and spleen datasets show that UNet 3+ outperforms previous state-of-the-art approaches, achieving higher Dice coefficients and producing more accurate, coherent segmentation results. The method is particularly effective for organs that appear at varying scales. The code is available at https://github.com/ZJUGiveLab/UNet-Version.