Noise-robust neural networks for medical image segmentation by dual-strategy sample selection

Deep neural networks for medical image segmentation often face the problem of insufficient clean labeled data. Although non-expert annotations are more readily accessible, these low-quality annotations lead to significant performance degradation of existing neural network methods. In this paper, we focus on robust learning of medical image segmentation with noisy annotations and propose a novel noise-tolerant framework based on dual-strategy sample selection, which selects the informative samples to provide effective supervision information. First, we propose the first round of sample selection by designing a novel joint loss, which includes conventional supervised loss and regularization loss. To further select information-rich samples, we propose confidence-based pseudo-label sample selection from a novel perspective as the complement. The dual strategies are used in a collaborative manner and the network is optimized with mined informative samples. We conducted extensive experiments on datasets with both simulated noisy labels and real-world noisy labels. For instance, on a simulated dataset with 25% noise ratio, our method achieves segmentation Dice value with 90.56% $\pm$ 0.03%. Furthermore, increasing the noise ratio to 95%, our method still maintains a high Dice value of 73.85% $\pm$ 0.28% compared to other baselines. Extensive results have demonstrated that our method can weaken the effects of noisy labels on medical image segmentation.