Deep Learning-Based COPD Detection: Neural Network Performance Comparison and Model Interpretability Study

Le résumé fourni par la source

COPD is a leading cause of death and often remains undiagnosed. Early detection through chest CT scans is crucial. This study tested several neural network models for detecting COPD using data from 3,599 patients. ResNet18 had the highest accuracy and sensitivity, while ResNet50 had the highest specificity. Adding an attention module to ResNet18 improved model interpretability, recommending ResNet18 for COPD screening. erj;66/suppl_69/PA979/TB1 T1 TB1 Table 1. Demographic and Clinical Characteristics of Participants Total (N=3599) COPD Group (N=1795) Control Group (N=1804) p-value Age (years) Mean ± SD 62.3±10.5 65.1±9.8 59.5±10.2 <0.001 Median (IQR) 63 (55-70) 66 (58-72) 60 (52-67) <0.001 Sex, n (%) <0.001 Male 2167 1121 1046 Female 1432 674 758 FEV1/FVC Mean ± SD 0.68±0.12 0.58±0.08 0.78±0.06 <0.001 erj;66/suppl_69/PA979/TB2 T2 TB2 Table 2: Performance Metrics of Seven Neural Network Architectures for COPD Detection Validation set (%) Test set (%) Accuracy Sensitivity Specificity AUC Accuracy Sensitivity Specificity AUC DenseNet 121 80.7 81.1 80.3 85.5 80.6 84.8 76.1 87.8 DenseNet 169 81.1 75.5 86.7 85.3 79.7 78.3 81.3 87.3 DenseNet 201 80.1 78.0 82.2 86.4 81.1 78.8 83.5 87.9 ResNet 18 82.2 86.4 78.1 87.3 83.3 89.1 77.3 88.9 ResNet 50 80.3 77.2 83.3 86.3 83.3 80.4 86.4 88.0 ResNet 152 80.8 78.6 83.1 87.3 82.2 79.9 84.7 89.0 Vision Transformer 76.5 72.4 80.6 80.7 73.9 70.7 77.3 81.5 erj;66/suppl_69/PA979/F1 F1 F1

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.