Globally, pneumonia is still a major health concern, particularly in areas with poor diagnostic facilities. This paper compares two CNN architectures, ConvXNet and a CustomCNN, for deep learning-based pneumonia identification using chest X-ray pictures. Preprocessing methods were used, including data augmentation, contrast enhancement, normalization, and grayscale conversion. A segmentation framework based on U-Net and ResNet32 was also implemented in order to separate lung regions and extract information unique to each region.

CustomCNN demonstrated strong generalization capabilities with a high training accuracy of 96.04%, while ConvXNet excelled in validation and test performance, achieving 88.94% validation accuracy and 90.75% test accuracy. Notably, CustomCNN showcased superior recall 98.5%, making it highly effective in minimizing missed pneumonia cases, whereas ConvXNet achieved slightly better precision 86.4%, ensuring fewer false positives. These findings highlight the complementary strengths of both architectures, emphasizing their potential in supporting accurate and reliable pneumonia detection and severity classification, especially in resource-constrained healthcare settings.