Pneumonia is a common and potentially lifethreatening lung infection, and chest radiography is usually the first test used to detect it. Many published deep learning results for this task are difficult to trust, not because the models are weak, but because of how they are evaluated. Recurring problems include model selection on a validation set of only sixteen images, reporting a single training run, relying on plain accuracy although normal and pneumonia cases appear in very unequal numbers, and tuning the decision threshold on the same images used for the final test. Rather than proposing a new architecture, this study defines and executes a reproducible evaluation protocol on the public pediatric chest X-ray benchmark. We reconstruct a validation split that preserves the original class balance, train every model with five random seeds and report the mean and spread of each metric, select the decision threshold on validation data only, and report specificity and balanced accuracy alongside recall, F1 score, and AUC. Three models are compared: a finetuned DenseNet-121, a fine-tuned VGG-16, and a small CNN trained from scratch. DenseNet-121 performs best on every metric, reaching an AUC of 0.963 ± 0.005 and a balanced accuracy of 0.802, with the most stable results across seeds. All three models detect almost every pneumonia case (recall ≈ 0.98) but correctly clear only 55–62% of normal radiographs, a weakness that plain accuracy hides entirely. Data splits, predictions, and checkpoints are released for full reproducibility.

Pneumonia detection, chest X-ray, deep learning, convolutional neural networks, transfer learning, reproducibility, class imbalance, Grad-CAM