The Value of Deep Learing for COPD Diagnosis and Grading

<bold>Background and Objectives</bold> Early diagnosis of Chronic Obstructive Pulmonary Disease (COPD) pose significant challenges. This study aims to develop a deep learning system for COPD detection algorithm and severity grading based on Computed Tomography (CT) imaging. <bold>Methods</bold> The patients with complete lung function and CT images were prospectively enrolled from the First Affiliated Hospital of Chongqing Medical University. We applied preprocessing with CT imaging though a window width of 124 and a window level of -962 to optimize feature extraction from low-attenuation areas. Patch-based sampling was employed to extract 125 patches of size 60x60x60 from lung regions with a stride of 60. These patches were input into a shared-parameter ResNet18 network for feature extraction and fusion, and fed into a fully connected layer for COPD detection or severity classification (GOLD 0-1 and GOLD 2-4 classification). The sensitivity and specificity of the predicting results were analyzed. <bold>Results</bold> The COPD detection algorithm and severity grading system was developed based on the 584 patients enrolled. In the COPD diagnosis task, a dataset of 584 images achieved an AUC of 0.927 on a test set of 40 images. For the mild (GOLD 0-1) and severe (GOLD 2-4) classification task, a dataset of 180 images attained a macro avg F1 score of 0.774 on a test set of 38 images. <bold>Conclusion</bold> Deep learning based on chest CT imaging holds enormous potential for COPD diagnosis and grading. <fig><object-id>erj;64/suppl_68/PA5240/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig>