AUTHOR=Cheng Xuan , Tian Haoxiang , Zhou Jiajun , Xie Tianshu , Gong HaiGang , Liu Ming , Wei Yi , Lu Wei 

TITLE=MPVT+: a noise-robust training framework for automatic liver tumor segmentation with noisy labels

JOURNAL=Frontiers in Medicine

VOLUME=Volume 12 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2025.1653865

DOI=10.3389/fmed.2025.1653865

ISSN=2296-858X

ABSTRACT=IntroductionLiver cancer is among the leading causes of cancer-related deaths worldwide. Accurate delineation of hepatic tumors is crucial for diagnosis, prognosis, and treatment planning, yet manual annotation is labor-intensive and subject to variability. Deep neural networks (DNNs) have shown promise in automating segmentation but require large amounts of high-quality labeled data, which is difficult to obtain. Incorporating noisy labels without proper handling can corrupt training and degrade performance.MethodsWe introduce MPVT+, a noise-robust training framework that integrates a pixel-wise noise-adaptation module with a multi-stage perturbation and variable-teacher (MPVT) consistency strategy. The noise adaptor infers corruption probabilities and re-weights unreliable supervision, while MPVT assembles an ensemble of stochastic teacher models that apply progressively stringent perturbations. This combination enables the network to exploit both clean and noisy labels without overfitting.ResultsExperiments conducted on 739 retrospectively collected liver-tumor CT datasets demonstrated that MPVT+ significantly outperformed baseline and traditional noise-handling approaches. Compared to a noise-free U-Net baseline (Dice Similarity Coefficient [DSC] 75.1%), MPVT+ improved segmentation accuracy to 80.3%. The framework consistently achieved superior results across multiple evaluation metrics, including DSC, JSC, SVD, and VOE.DiscussionThe MPVT+ framework demonstrates that principled noise modeling, coupled with consistency training, effectively unlocks the potential of imperfect medical datasets. This strategy reduces the dependency on perfectly labeled datasets and moves fully automated liver tumor delineation closer to clinical applicability.