A Deep Learning Method for Pneumoconiosis Staging on Chest X-Ray Under Label Noise

The ambiguous properties of small opacities in pneumoconiosis chest radiographs can cause diagnostic drift, which in turn leads to the presence of noisy labels in the datasets collected from hospitals that can negatively impact the generalization of deep learning models. To tackle this issue, we propose COFINE, a novel coarse-to-fine noise-tolerant deep learning method for the staging of pneumoconiosis chest radiographs, which comprises two procedures: coarse screening and fine learning. In the coarse screening procedure, the proposed sample selection strategy divides the pneumoconiosis dataset into ‘confident’ and ‘less-confident’ subsets based on the logical relationship between the prediction correctness and confidence of multiple expert networks. During the fine learning procedure, we apply two different strategies to fit the sample features in the above two subsets. For the samples in ‘confident’ subset, we specifically design a novel soft label relaxation learning strategy (SLRL) to mine the explicit features and overcome the overfitting problem caused by traditional one-hot labels. To parse the implicit features within the less-confident samples, the augmentation-based self-supervised learning method is employed. On the pneumoconiosis dataset containing 1,372 chest X-rays collected from West China Fourth Hospital, the proposed method achieves a sensitivity of 82.4% in identifying ‘stage-1’ cases, which is higher than that of other algorithms (70.6%), and is especially beneficial for clinical radiologists in pneumoconiosis staging.