Dual-Perspective Label Smoothing for Accurate Tumor Segmentation

Lilong Wang, Xuan Wang, Yali Xu, Xiangan Wu, Shunda Du, Shaoting Zhang, Qiang Sun, Xiaosong Wang

Abstract: Accurately segmenting tumors from medical images is crucial for cancer diagnosis and quantitative assessment. However, pixel-wise ground truth labeling is often over-confident when used directly as the training objective for deep learningbased image segmentation models due to factors like ambiguous tumor boundaries and inter-observer variability. To address this issue, we introduce dual-perspective soft labels for accurate tumor segmentation. First, it leverages inherent information from the training images and annotations, such as intensity and structure features, to formulate fused soft labels. Additionally, we explore a new type of soft label that is distilled from the model prediction (as the training feedback) in an iterative refinement manner, as well as adaptively adjusts the weights of the training samples. The proposed method demonstrates significant improvements in tumor segmentation accuracy across several popular segmentation backbone networks. It outperforms both baselines and existing label smoothing methods in extensive experiments conducted on tumor segmentation tasks in CT liver, CT kidney, and MR brain volumetric images. Code is available at https://github.com/wllfore/DPLS.