What is it about?
Pneumonia causes around 14% of all deaths in children under five. Chest X-rays are the standard way to diagnose it, but reading them takes an experienced radiologist — and in many hospitals, particularly in low-resource settings, there simply aren't enough. We built and tested an artificial intelligence system that examines a chest X-ray and estimates the probability that the patient has pneumonia. Rather than using one of the very large models common in this field, we deliberately chose a compact one, EfficientNetV2B0, with about 7.1 million parameters — small enough to run on ordinary hardware. We trained it in two stages on 5,863 children's X-rays from the widely used Kermany dataset, first adapting only the final decision layers and then fine-tuning the whole network. The system correctly identified 98% of pneumonia cases, with an overall accuracy of 89.4% and an AUC of 0.967. It is more cautious in the other direction: about a quarter of healthy children were flagged for a second look. We also applied a technique called Grad-CAM, which produces a heat map showing which parts of the X-ray drove each decision — confirming the model was looking at genuine lung abnormalities rather than incidental image features. Every experiment was run with fixed random seeds on a single CPU so that anyone can reproduce our numbers exactly.
Featured Image
Photo by Bozhin Karaivanov on Unsplash
Why is it important?
In medicine, not all mistakes cost the same. Missing a child with pneumonia can be fatal; wrongly flagging a healthy child means an unnecessary follow-up. Our model is tuned toward the safer error — it catches 98% of real cases, meaning roughly 2% would be missed, while over-referring some healthy patients. For triage and screening, that is the right trade-off. The practical contribution is that this level of performance does not require a large model. Much of the published work reaching comparable accuracy uses architectures like DenseNet-121 or EfficientNetV2-XL, which need substantial computing power and reliable bandwidth. A 7.1-million-parameter model can run on edge devices or embedded systems, making it realistic for rural clinics and telehealth networks where the need is greatest. Equally important is how the work was done. We followed the CLAIM and TRIPOD-AI reporting standards, disabled GPU computation to guarantee deterministic results, reported bootstrapped confidence intervals across 1,000 resamples, and tested statistical significance with the DeLong test. Reproducibility is a persistent weakness in medical AI research; a result that cannot be independently verified cannot responsibly be brought near a patient.
Perspectives
The finding we did not expect was in the calibration analysis. The model is accurate, but when it outputs a low probability — say 0.1 to 0.4 — actual pneumonia occurs more often than that number suggests. It is systematically under-confident in exactly the range where a clinician might be tempted to discharge a patient. Accuracy alone would never have revealed this; only measuring calibration did, and we report an Expected Calibration Error of 0.083 and a Brier Score of 0.091 so others can see it too. Post-hoc methods such as temperature scaling or isotonic regression should close the gap, and we think reporting calibration should be standard rather than optional in clinical AI papers. We are also clear about the limits. The Kermany dataset comes from a single medical centre in Guangzhou and contains only paediatric radiographs. Strong numbers on one dataset from one scanner and one population are not evidence a system will work elsewhere. Cross-dataset validation against ChestX-ray14, CheXpert, or the RSNA challenge data is the necessary next step, and no deployment claim should be made before it.
Assoc. Prof. Narasimha Rao Vajjhala
University of New York at Tirana
Read the Original
This page is a summary of: Deep Learning for Medical Image Analysis: CNN-based Pneumonia Detection on Chest X-Rays, May 2026, Institute of Electrical & Electronics Engineers (IEEE),
DOI: 10.1109/icpcsn68523.2026.11543623.
You can read the full text:
Contributors
The following have contributed to this page
