Adaptive Cross-Attention for Robust Lung Segmentation with Noisy Labels

Accurate lung segmentation in computed tomography (CT) images is essential for diagnosing and managing respiratory diseases. However, achieving reliable segmentation is challenging due to the scarcity and noise of labeled medical data. In this paper, we present a novel approach for lung segmentation under noisy labeling conditions via attention-guided transfer learning. Our method integrates an Adaptive Cross-Attention block, which selectively incorporates features from pretrained models to improve robustness in segmentation. Furthermore, we introduce an adaptive loss function that dynamically weighs pretrained and ground-truth outputs during training, effectively mitigating the impact of possible label noise. Extensive evaluations on public datasets demonstrate that our approach improves segmentation accuracy, achieving an average Dice coefficient increase of 3.7% on clean data and a significant 29.4% improvement under simulated noisy annotations compared to the standard nnU-Net. These results establish our model as a simple and effective solution for lung segmentation under label noise constraints.