CT-based COPD identification using multiple instance learning with two-stage attention

Chronic obstructive pulmonary disease (COPD) is one of the leading causes of morbidity and mortality worldwide. However, COPD remains underdiagnosed globally. Spirometry is currently the primary tool for diagnosing COPD, but it has unneglected difficulties in detecting mild COPD. Chest computed tomography (CT) has been validated for COPD diagnosis and quantification. Whereas many CT-based deep learning approaches have been developed to identify COPD, it remains challenging to characterize CT-based pathological alternations of COPD which are multidimensional and highly spatially heterogeneous, and the diagnosis performance still needs to be improved.A multiple instance learning (MIL) with two-stage attention (TSA-MIL) is proposed to identify COPD using CT images. Based on transfer learning, a Resnet-50 model pre-trained on natural images is used to extract multicomponent and multidimensional features of COPD abnormalities, in which a pseudo-color method is designed to transfer single-channel CT slices to RGB-like three channels and meanwhile increase the richness of feature representations. To generate more robust attention score for each instance, a two-stage attention module is utilized with the first stage aiming at discovering the key instance while the second stage correcting the attention score for each instance by calculating its average relative distance to the key instances; besides, an instance-level clustering over feature domain is exploited to further improve feature separability and therefore facilitate the subsequent attention module. CT scans, spirometry and demographic data of a total of 800 participants were collected from a large public hospital, with 720 and 80 participants used for model development and evaluation, respectively. In addition, data of 260 participants from another large hospital were also collected for external validation.The proposed TSA-MIL approach outperforms not only most of the advanced MIL models, but also other up-to-date COPD identification methods, with an accuracy of 0.9200 and an area under curve (AUC) of 0.9544 on the test set, and with an accuracy of 0.8115 and an AUC of 0.8737 on the external validation set without multicenter effect reduction, which is clinically acceptable. Therefore, this approach is promising to be a powerful tool for COPD diagnosis in clinical practice.