Comparative Analysis of Robustness of ResNet-18 and ResNet-50 Architectures to Noise Distortions in Pneumonia Classification from Chest X-ray Images

We conducted a comparative analysis of two convolutional neural network architectures, ResNet-18 and ResNet-50, for diagnosing pneumonia from chest X-ray images. We trained the models using transfer learning on the Chest X-Ray Pneumonia dataset. We evaluated predictive performance (Accuracy, F1, AUC-ROC), computational efficiency, and robustness to Gaussian noise. We found that ResNet-50 outperforms ResNet-18 both on clean data (AUC = 0.975 vs. 0.940) and under noisy conditions: at a noise level of σ = 0.1, the AUC of ResNet-50 is 0.865, whereas ResNet-18 drops to 0.649. This indicates higher robustness of ResNet-50 to noise. Thus, ResNet-50 is the preferred architecture for deployment in medical institutions, especially when image quality is inconsistent.