• A Simple yet Effective Data Scaling Strategy for Semi-Supervised Medical Image Segmentation

    Liu, Yajun

    Abstract

    Semi-supervised medical image segmentation (SSMIS) aims to enhance segmentation performance by leveraging both limited labeled data and abundant unlabeled data, thereby reducing the reliance on costly manual annotations. Recent consistency-based approaches, particularly those involving image-level copy-paste augmentation, often suffer from anatomically implausible artifacts near pasted boundaries, which compromise the integrity of anatomical structures. To address this limitation, we propose a concise yet effective approach that relies solely on Data Scaling to preserve the anatomical integrity of segmentation structures while improving segmentation performance. Our Data Scaling strategy includes two parallel operations: Multi-Scale Random Scaling (MSRS) for labeled data and Difficulty-Driven Adaptive Scaling (DDAS) for unlabeled data. Specifically, MSRS expands the spatial diversity of labeled training samples through randomized multi-scale local and global scaling, effectively capturing local and global features. DDAS leverages prediction disagreement between dual networks to estimate pixel-level segmentation difficulty, adaptively scaling challenging regions to boost model learning effectiveness on unlabeled data. Experiments show that our method achieves state-of-the-art performance on 2D and 3D medical datasets.