Pediatric Pneumonia Chest X-ray Dataset

Updated 2 July 2026

Pediatric pneumonia CXR datasets are structured collections of annotated chest X-rays from patients aged 0–15 years, serving as a gold standard for automated diagnosis.
They incorporate expert-reviewed labeling, standardized preprocessing, and augmentation techniques to address class imbalance and improve model performance.
The datasets enable rigorous evaluations using metrics like accuracy, recall, and AUC, while supporting explainable AI methods for clinical validation.

Pediatric pneumonia chest X-ray (CXR) datasets are central to the development, benchmarking, and deployment of automated deep learning models for the diagnosis of lower respiratory tract infection in children. These datasets consist of CXR images—primarily acquired from pediatric patients aged 0–15 years—meticulously annotated by clinical experts for pneumonia and related pathologies. The most widely used repositories, such as the dataset introduced by Kermany et al. (2018), serve as a gold standard, referenced in numerous machine learning studies for algorithm validation, transfer learning, and cross-institutional research.

1. Major Datasets and Their Provenance

The dominant dataset in the literature is the Guangzhou Women and Children’s Medical Center (GWCMC) Pediatric Pneumonia Chest X-ray Dataset. This resource is available via Mendeley Data and Kaggle, comprising 5,856 pediatric CXR images—1,583 labeled as normal and 4,273 as pneumonia (further subdivided into 1,493 viral and 2,780 bacterial cases). Images were obtained from children aged 1–5 years under IRB-approved protocols, with ground-truth labels adjudicated by board-certified radiologists (Neshat et al., 2023). Further expert-reviewed chest X-ray datasets include:

PediCXR: 9,125 PA-view pediatric CXRs from Vietnam, with per-image labels for 15 diseases (including pneumonia), accompanying bounding boxes for 36 local findings (e.g., consolidation), and precise DICOM metadata (Pham et al., 2022).
VinDr-PCXR: >9,000 annotated pediatric CXRs from Vietnam, integrating both diagnosis and lesion-level information (Raghaw et al., 2024).
P2Med-MD (P2Med-MLLM): 67,616 pediatric CXR DICOM studies with multi-label taxonomy and severity scores, covering ages from neonate to 15 years with radiologist-verified annotations (Tian et al., 2024).
Derived and extended datasets: Additional repositories for specific clinical subtypes (e.g., Mycoplasma pneumoniae) or for multi-modal/federated learning contexts combine primary GWCMC images with regionally acquired or institutionally collected scans (Deng et al., 2024, Jimenez-Gutierrez et al., 12 Nov 2025).

2. Data Composition, Annotation, and Label Taxonomy

The standard GWCMC dataset's ground-truth pipeline involves dual radiologist readouts, consensus or adjudication on disagreement, and class assignments of “Normal,” “Bacterial Pneumonia,” or “Viral Pneumonia.” Labeling definitions are standardized: pneumonia cases show radiographic consolidation and/or infiltrate, while normals lack such findings (Neshat et al., 2023, Roeder, 2023). PediCXR and VinDr-PCXR further annotate bounding boxes for consolidation, pleural effusion, and other findings, and in some datasets, granular disease sub-labels (lobar, bronchopneumonia, etc.) are extracted semi-automatically or by manual curation (Pham et al., 2022, Tian et al., 2024).

Clinical metadata such as age, gender, physical size, or ancillary laboratory data is typically sparse in public datasets, although P2Med-MD and PediCXR include limited clinical or acquisition metadata in their DICOM headers and auxiliary files (Tian et al., 2024, Pham et al., 2022).

3. Image Characteristics and Preprocessing Protocols

Images originate as single-channel (grayscale) digital radiographs, resolutions ranging from 512×512 (GWCMC) up to 1600×1300 (PediCXR). Preprocessing aligns with neural network requirements:

Resizing: Images are interpolated to network-specific resolutions, commonly 224×224 or 299×299 px (Neshat et al., 2023, Chauhan et al., 13 Oct 2025, Ahmed, 10 Sep 2025).
Intensity normalization: Pixel intensities $X \in [0,255]$ are normalized via $X_\mathrm{scaled} = X/255$ or $X_\mathrm{norm} = (X-\mu)/\sigma$ where $\mu,\sigma$ are computed across the training cohort (Neshat et al., 2023).
Channel handling: Replication of grayscale images into three identical channels is standard for models pretrained on ImageNet (Chauhan et al., 13 Oct 2025, Ahmed, 10 Sep 2025).
Advanced preprocessing: Some cohorts undergo lung windowing, CLAHE (clip limit=2.0, grid=8×8), rib suppression, and segmentation to standardize image contrast and suppress anatomical confounders (Raghaw et al., 2024).

4. Splitting Strategies, Class Imbalance, and Augmentation

Standard practice splits datasets into train, validation, and test partitions, using stratified sampling to maintain proportions of normals and pneumonia (Neshat et al., 2023, Angara et al., 2024). The GWCMC dataset often follows an 80% train, 10% validation, 10% test convention, or as provided in the released folders (e.g., train = 5,216, test = 624, val = 16) (Chauhan et al., 13 Oct 2025). Class ratios are highly imbalanced, frequently ∼3:1 in favor of pneumonia. Data augmentation is performed exclusively on training data, leveraging random horizontal flips (p=0.5), rotations (e.g., $\theta \sim U(-15^\circ,15^\circ)$ ), shift, shear, zoom, and intensity jitter (Neshat et al., 2023, Manaf et al., 13 Jul 2025). Some works synthesize additional minority-class images with GANs or adversarial augmentation to mitigate imbalance (Manaf et al., 13 Jul 2025, Raghaw et al., 2024).

5. Annotation/Lesion Localization and Explainability

Class labels—binary pneumonia vs. normal and, less frequently, bacterial vs. viral—are established via radiologist consensus (typically at least two out of three experts), with QA to remove ambiguous or low-quality images (Roeder, 2023). Certain datasets provide lesion-level bounding boxes for “consolidation” or “opacity,” e.g., PediCXR, enabling weakly supervised learning and precise explainability (Pham et al., 2022). Explainable AI methods such as Grad-CAM, Score-CAM, and LIME are routinely integrated to localize decision regions and visualize salient features within the lung fields, supporting clinical verification and algorithmic audit (Deng et al., 2024, Khadidos et al., 14 Jan 2026).

6. Benchmarking, Evaluation Metrics, and Comparative Performance

Evaluation protocols use well-accepted medical imaging metrics. In all cited works, let TP, TN, FP, FN denote true/false positives/negatives, with the positive class defined as “pneumonia.” The primary metrics are:

Accuracy: $\mathrm{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}$
Precision (PPV): $\frac{TP}{TP + FP}$
Recall (Sensitivity, TPR): $\frac{TP}{TP + FN}$
Specificity (TNR): $\frac{TN}{TN + FP}$
F1-score: $F1 = 2 \frac{\mathrm{Precision} \times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$
Matthews Correlation Coefficient (MCC): $X_\mathrm{scaled} = X/255$ 0 (Khadidos et al., 14 Jan 2026).

Recent advanced architectures (e.g., Inception-ResNet-V2, RepViT-CXR, LightPneumoNet) achieve test-set accuracies and AUC in the range $X_\mathrm{scaled} = X/255$ 1– $X_\mathrm{scaled} = X/255$ 2, with the best ViT-based models exceeding $X_\mathrm{scaled} = X/255$ 3 accuracy, $X_\mathrm{scaled} = X/255$ 4 recall, and $X_\mathrm{scaled} = X/255$ 5 AUC (Neshat et al., 2023, Ahmed, 10 Sep 2025, Chauhan et al., 13 Oct 2025). Classifier comparisons and ablation studies demonstrate superior feature extraction and computational efficiency for hybrid, transformer, and lightweight custom CNN models under identical dataset and preprocessing constraints (Neshat et al., 2023).

7. Regulatory Compliance, Access, Limitations, and Recommendations

The GWCMC dataset and its derivatives are available for academic use under Mendeley Data and Kaggle Terms, with anonymized DICOM/JPEG/PNG files and explicit license statements (Neshat et al., 2023, Chauhan et al., 13 Oct 2025). PediCXR and VinDr-PCXR are available through PhysioNet under credentialed agreements (Pham et al., 2022, Raghaw et al., 2024). Federated and privacy-preserving scenarios simulate multi-institutional partitions and non-IID data using Dirichlet-based splits, achieving high accuracy without data centralization and meeting HIPAA/GDPR requirements (Jimenez-Gutierrez et al., 12 Nov 2025).

Limitations of canonical datasets include single-center source, lack of large-scale diverse metadata (e.g., gender, comorbidities), class imbalance, patient-level repetition, and restriction to a narrow pediatric age range (typically 1–5 years). Recommendations for robust downstream research include curating multicenter, demographically balanced, and multi-modal CXR datasets, reporting inter-rater variability, applying class balancing or weighting in model training, and validating algorithms on external test sets for generalizability (Roeder, 2023, Neshat et al., 2023).

References:

(Neshat et al., 2023, Angara et al., 2024, Chauhan et al., 13 Oct 2025, Ahmed, 10 Sep 2025, Jimenez-Gutierrez et al., 12 Nov 2025, Raghaw et al., 2024, Deng et al., 2024, Roeder, 2023, Hosseinabadi et al., 31 Dec 2025, Pham et al., 2022, Tian et al., 2024, Liz et al., 2020, Manaf et al., 13 Jul 2025, Khadidos et al., 14 Jan 2026).