Deep Learning for Medical Image Analysis: CNN-based Pneumonia Detection on Chest X-Rays

Summary

What question does this paper answer?

Can a lightweight EfficientNetV2B0 convolutional neural network, trained with transfer learning under strict reproducibility protocols, accurately and reliably detect pneumonia in paediatric chest X-rays?

What did the study find?

On the Kermany paediatric chest X-ray dataset (624-image test set), the EfficientNetV2B0 model achieved an AUC of 0.967 (95% CI 0.953–0.979; DeLong p < 0.001), accuracy of 0.894, F1-score of 0.920 and pneumonia recall of 0.98, missing only about 2% of true pneumonia cases. Calibration was moderate (ECE = 0.083, Brier Score = 0.091), and normal-class recall was lower at 0.752.

Why does it matter?

The results show that a compact model of about 7.1 million parameters can deliver clinically meaningful pneumonia detection without the heavy computation of larger networks, making it suited to edge devices, telehealth networks and low-resource healthcare settings. Its high sensitivity favours triage workflows, where missing pneumonia is costlier than an unnecessary follow-up, and the authors outline integration with PACS and CMMS platforms.

Study at a glance

Abstract

Pneumonia remains one of the leading causes of mortality worldwide, particularly among paediatric populations, and timely radiographic diagnosis is essential for effective clinical intervention. The proliferation of pervasive computing and communication technologies has enabled the deployment of deep learning models within distributed healthcare systems, facilitating rapid and automated medical image analysis at scale. This study presents a convolutional neural network framework based on EfficientNetV2B0 with transfer learning for binary pneumonia classification from paediatric chest X-ray images. The model employs a two-phase training strategy comprising frozen-backbone feature extraction followed by full fine-tuning, augmented by deterministic data preprocessing, class-weighted loss computation, and Gradient-weighted Class Activation Mapping for interpretability. The input to the model consists of 224×224 pixel RGB chest radiographs, and the output is a sigmoid-activated probability score indicating the likelihood of pneumonia. Evaluated on the benchmark Kermany dataset of 5,863 radiographs (1,583 normal and 4,280 pneumonia cases across training, validation, and test splits), the proposed framework achieves an area under the receiver operating characteristic curve (AUC) of 0.967, confirmed as statistically significant via DeLong testing (p < 0.001), and an overall accuracy of 0.894, with a pneumonia recall of 0.98. The model further yields an Expected Calibration Error (ECE) of 0.083 and a Brier Score of 0.091. Bootstrapped confidence intervals and calibration analysis confirm the robustness and reliability of the predictions. The results demonstrate that lightweight EfficientNetV2 architectures, when combined with rigorous reproducibility protocols compliant with CLAIM and TRIPOD-AI standards, can deliver clinically meaningful diagnostic performance suitable for integration into pervasive healthcare information and communication technology platforms.

Abstract as published in 2026 6th International Conference on Pervasive Computing and Social Networking (ICPCSN).

Keywords: deep learning; convolutional neural networks; pneumonia detection; chest X-ray classification; EfficientNetV2; transfer learning; medical image analysis; pervasive computing

Key terms

Transfer learning

Adapting a model pretrained on a large dataset (here ImageNet) to a specific task such as chest X-ray classification, useful when annotated clinical data are limited.

Grad-CAM (Gradient-weighted Class Activation Mapping)

A technique that produces heatmaps highlighting the image regions most influential in a model’s classification decision.

Expected Calibration Error (ECE)

The weighted average absolute difference between a model’s predicted confidence and its observed accuracy across probability bins.

Limitations

• The Kermany dataset contains only paediatric chest radiographs from a single medical centre, limiting demographic and scanner diversity and external validity; cross-dataset evaluation (e.g., NIH ChestX-ray14, RSNA, CheXpert) is needed.
• The model was trained on a relatively small dataset, so performance consistency on larger, more heterogeneous datasets remains to be verified.
• Calibration analysis shows systematic under-confidence in low-probability regions (ECE = 0.083), so probabilities may need post-hoc calibration before clinical risk stratification.
• Grad-CAM captures only high-level activation patterns and does not guarantee clinically valid explanations.

How to cite

Haveri, K., & Vajjhala, N. R. (2026). Deep Learning for Medical Image Analysis: CNN-based Pneumonia Detection on Chest X-Rays. In 2026 6th International Conference on Pervasive Computing and Social Networking (ICPCSN) (pp. 156–161). IEEE. https://doi.org/10.1109/ICPCSN68523.2026.11543623

@inproceedings{haveri2026cnn,
  title = {Deep Learning for Medical Image Analysis: CNN-based Pneumonia Detection on Chest X-Rays},
  author = {Haveri, Katia and Vajjhala, Narasimha Rao},
  booktitle = {2026 6th International Conference on Pervasive Computing and Social Networking (ICPCSN)},
  pages = {156--161},
  year = {2026},
  publisher = {IEEE},
  doi = {10.1109/ICPCSN68523.2026.11543623},
  url = {https://doi.org/10.1109/ICPCSN68523.2026.11543623}
}