Pulmonary Segments Segmentation with Hierarchical Weak Labels.

Accurate segmentation of pulmonary segments in computerized tomography (CT) images is crucial for the quantitative analysis of lung diseases, such as the lung cancer and coronavirus disease. However, it is prohibitive to train a network in a fully supervised manner for two reasons: precise labeling requires experienced annotators, and complete labeling is time-consuming. In light of this, we propose a novel end-to-end training strategy which exploits weakly labeled data. Based on the pulmonary hierarchical anatomy (the fact that segments compose lobes and lobes compose the lung), a hierarchical loss enhances the supervision information with additional full lobe labels and lung labels. Besides, a continuity loss is proposed to encourage continuous segmentation and smooth boundaries. Experiments on three datasets demonstrate that the proposed method outperforms previous state-of-the-art methods for segment, lobe and lung segmentation, respectively. Furthermore, pulmonary nodules can be accurately localized provided the segment segmentation result, where our method achieves 12% advance on accuracy in comparison with other methods.