Adaptive Multi-Scale Local Correction for Semi-Supervised 3D Medical Image Segmentation

In recent years, semi-supervised 3D medical image segmentation has gained significant attention. However, current methods often struggle with multi-scale voxel differences and overlook the importance of loss weight balancing. To address these issues, we propose an adaptive multi-scale local correction method (AMLC). Our key contributions are: (1) a multi-scale local correction module to accurately capture differences between 3D voxel blocks at various scales; (2) an adaptive weighting adjustment module that dynamically balances losses, improving robustness and training. Experiments show that AMLC significantly outperforms existing methods on the public dataset, demonstrating its effectiveness.