MPVT+: A Noise-robust Training Framework for Automatic Liver Tumor Segmentation with Noisy Labels

Xuan Cheng¹Haoxiang Tian²Jiajun Zhou²Tianshu Xie¹Haigang Gong³Ming Liu³Yi Wei^4*Wei Lu^3*

• ¹University of Electronic Science and Technology of China, Chengdu, China
• ²University of Electronic Science and Technology of China Yangtze Delta Region Institute Quzhou, Quzhou, China
• ³People's Hospital of Quzhou, Quzhou, China
• ⁴West China Hospital of Sichuan University, Chengdu, China

Liver cancer is a leading cause of cancer-related deaths globally. Accurate delineation of hepatic tumours underpins diagnosis, prognostication and treatment planning, yet manual annotation remains onerous. Significant progress in segmentation has been realized through the utilization of deep neural networks (DNNs). But the training process of deep learning requires a substantial amount of data with high-quality labels, which is particularly challenging to obtain. Simply incorporating the more abundant, yet error-prone, annotations can instead corrupt learning and potentially lead to suboptimal segmentation performance and decreased generalization ability. Here we introduce MPVT+, a noise-robust training paradigm that couples a label-noise adaptor with a multi-stage perturbation and variable-teacher (MPVT) consistency framework. The adaptor infers pixel-wise corruption probabilities and dynamically re-weights unreliable supervision. In parallel, the MPVT module assembles an ensemble of stochastic "teacher" networks that expose the student to progressively stringent perturbations. This synergy allows the network to exploit imperfect data while resisting over-fitting. Experimental results demonstrate that the MPVT+ framework outperforms traditional methods in liver tumor segmentation accuracy. Our findings show that principled noise modelling, paired with consistency training, can unlock the latent value of imperfect medical datasets and move fully automated liver-tumour delineation closer to routine clinical practice.