Effect of Computed Tomography Window Technique on the Results of Artificial Intelligence Classification of Lung Lesions

Objective: To use three different 3D CNN algorithms and five different computed tomography (CT) window settings to study the effect on the results of artificial intelligence classification of lung lesions in different CT window techniques. Method: A total of 172 cases of peripheral lung cancer and 185 of focal pneumonia who underwent chest CT were analyzed. Three different 3D CNN algorithms were selected (ResNet, ResNext, and DenseNet) to divide the lesions into two groups. Five different CT window settings, including lung window (1500, 600), mediastinal window (350, 40), custom window 1 (SW1) (1000, 40), and custom window 2 (SW2) (1000, 100), were used retrospectively. We calculated classification accuracy, receiver operating characteristic (ROC) curve, and area under the curve (AUC). The ROC curve was compared in pairs. Results: The average classification accuracy of ResNet was the lowest in the mediastinal window (85.732%; AUC value: 0.871) and the highest in the full window (91.596%; AUC value: 0.946). The average classification accuracy of ResNext was the lowest in the mediastinal window (81.528%; AUC value: 0.814) and the highest in the full window (86.568%; AUC value: 0.882). The average classification accuracy of DenseNet was the lowest in the mediastinal window (87.954%; AUC value: 0.906) and the highest in the SW2 window (93.274%; AUC value: 0.951). Medcalc was used to compare ROC curves under five windows of three 3D CNN. The AUC values between mediastinal window and lung window, mediastinal window and SW1, and mediastinal window and SW2 were statistically significant. Conclusion: There is little difference in the diagnostic efficacy of the three 3D CNN. Different CT window settings have an influence on the results of CNN classification of the lung lesions, and the diagnostic efficiency of the three 3D CNN is the worst under the mediastinal window.