Learning to Segment from Noisy Annotations: A Spatial Correction Approach
This addresses the challenge of error-prone annotations in medical imaging, which can degrade segmentation performance, by providing a more accurate approach for handling spatially correlated noise.
The paper tackles the problem of noisy labels in medical image segmentation by proposing a novel Markov model that accounts for spatial correlation and bias, and a label correction method, showing it outperforms state-of-the-art methods on synthetic and real-world datasets.
Noisy labels can significantly affect the performance of deep neural networks (DNNs). In medical image segmentation tasks, annotations are error-prone due to the high demand in annotation time and in the annotators' expertise. Existing methods mostly assume noisy labels in different pixels are \textit{i.i.d}. However, segmentation label noise usually has strong spatial correlation and has prominent bias in distribution. In this paper, we propose a novel Markov model for segmentation noisy annotations that encodes both spatial correlation and bias. Further, to mitigate such label noise, we propose a label correction method to recover true label progressively. We provide theoretical guarantees of the correctness of the proposed method. Experiments show that our approach outperforms current state-of-the-art methods on both synthetic and real-world noisy annotations.