Pediatric Pneumonia Classification System

• Pediatric chest pneumonia classification systems are computational frameworks that analyze pediatric radiographs using deep learning and multimodal methods to support clinical decision-making.
• They employ diverse architectures including custom CNNs, transfer learning models, and Vision Transformers to differentiate between normal and pneumonia cases, as well as subtype distinctions.
• Key challenges include data imbalance, annotation variability, and the need for external validation to ensure clinical reliability and effective deployment.

A Pediatric Chest Pneumonia Classification System is a computational decision-support framework for analyzing pediatric chest radiographs, most often to classify images as Normal or Pneumonia, but in some studies to distinguish bacterial, viral, mycoplasma, or other pediatric thoracic categories. In the recent literature, such systems span custom CNNs, transfer-learned CNNs and Vision Transformers, ensemble models, anomaly-detection pipelines, federated-learning deployments, explainable AI overlays, and multimodal assistants that combine radiographs with symptoms, cough, speech, or longitudinal clinical records. The topic is pediatric-specific because several papers explicitly emphasize smaller thoracic anatomy, less distinct pulmonary patterns than adults, motion and cooperation difficulties during image acquisition, and the clinical importance of rapid differentiation in children, especially under five years of age (Manaf et al., 13 Jul 2025, Hosseinabadi et al., 31 Dec 2025).

1. Clinical scope and task formulations

The dominant formulation is binary image-level classification of pediatric chest radiographs into Normal versus Pneumonia. This is the explicit target in the CNN-plus-GAN pipeline, the transfer-learning comparison of ResNetRS/RegNet/EfficientNetV2, the ResNet-50 explainability study, the lightweight LightPneumoNet model, the federated ResNet18 study, the CNN–Vision Transformer fusion model, and the ViT channel-replication study (Manaf et al., 13 Jul 2025, Hosseinabadi et al., 31 Dec 2025, Ridwan, 18 Jul 2025, Chauhan et al., 13 Oct 2025, Jimenez-Gutierrez et al., 12 Nov 2025, Angara et al., 2024, Ahmed, 10 Sep 2025). In these works, the intended role is usually decision support, screening, triage, or a second-reader function rather than autonomous diagnosis.

A second family of systems addresses etiologic or subtype discrimination. One paper decomposes the problem into two binary tasks, first normal vs pneumonia and then viral vs bacterial pneumonia, using transfer-learned CNNs, optional feature selection, and SVM classification (Arizmendi et al., 3 Mar 2025). FA-Net evaluates both binary and 3-class setups, with the multi-class problem defined as bacterial pneumonia, viral pneumonia, and normal (Roy et al., 2024). “Pneumonia App” extends the task further to four classes—Normal, Bacterial pneumonia, Viral pneumonia, and Mycoplasma pneumoniae pneumonia (MPP)—for children aged 0–12 years (Deng et al., 2024). A different clinical reframing appears in the XrPP ensemble study, where the classification target is consolidation (alveolar pneumonia) versus non-consolidation (other infiltrates / non-alveolar pneumonia) rather than normal-vs-disease screening (Liz et al., 2020).

A third formulation expands beyond pure radiographic classification. MultiSense-Pneumo maps symptom descriptors, cough audio, spoken language, and chest radiographs into normalized risk signals and fuses them into a triage score $S \in [0,1]$, which is then converted into LOW / MODERATE / HIGH risk bands (Jayakody et al., 4 May 2026). P2Med-MLLM is broader still: it is a multimodal generative assistant for pediatric pneumonia-centered care that produces free-text radiology reports, outpatient records, disease-course notes, and ward-round records rather than a narrow discriminative class label (Tian et al., 2024). These extensions indicate that “classification system” in this domain now encompasses not only image-level prediction but also workflow-aware assistance, longitudinal reasoning, and multimodal triage.

2. Data foundations and annotation regimes

Two data resources dominate the literature. The first is the Guangzhou/Kermany pediatric chest X-ray corpus, distributed through Kaggle or Mendeley and reused across many papers. The second is PediCXR, an open pediatric chest radiograph dataset with image-level disease labels and lesion-level bounding boxes (Pham et al., 2022).

Resource	Reported properties	Relevance
Guangzhou/Kermany pediatric CXR dataset	Reported as 5,856, 5,863, or 5,864 images; ages reported as 0–5 or 1–5 years; binary Normal/Pneumonia or subtype labels	Main benchmark for binary pediatric pneumonia classification
PediCXR	9,125 studies, one image per patient, children younger than 10 years, 36 local findings and 15 diseases plus No finding	Supports pediatric disease classification plus lesion-level localization
XrPP	950 PA pediatric radiographs, ages 1 month–16 years, labels: consolidation vs non-consolidation	Supports radiographic-pattern classification rather than only normal-vs-pneumonia

Across papers, the Guangzhou/Kermany dataset is not reported uniformly. One study describes 5,863 chest X-ray images from children aged 0 to 5 years, labeled Normal and Pneumonia, with labels assigned by two physicians and verified by a third-party evaluator (Manaf et al., 13 Jul 2025). Another reports 5,856 anterior-posterior chest radiographs from pediatric patients aged 1 to 5 years (Hosseinabadi et al., 31 Dec 2025). A separate transfer-learning study states 5864 grayscale chest X-ray images with 2580 bacterial pneumonia, 1500 viral pneumonia, and 1784 normal, then downsamples to 1500 images per class (Arizmendi et al., 3 Mar 2025). The ResNet-50 explainability paper reports 5,863 grayscale anterior-posterior radiographs, then states that 7 images were excluded due to file corruption or unreadable format, leaving 5,856 images actually used (Ridwan, 18 Jul 2025). This reporting variation is itself a recurrent reproducibility issue.

PediCXR provides a different annotation regime and task space. It contains 9,125 studies/images from 9,125 patients, split into 7,728 training and 1,397 test studies, with each study consisting of a single posteroanterior (PA) view chest radiograph in DICOM format (Pham et al., 2022). The paper presents PediCXR as the first and largest public pediatric CXR dataset with both image-level and lesion-level annotations, including 36 local finding labels and 15 diseases plus “No finding.” Pneumonia appears explicitly as a disease label, alongside Brocho-pneumonia and Pleuro-pneumonia, while pneumonia-associated findings such as Consolidation, Infiltration, Pleural effusion, and Diffuse aveolar opacity are separately box-annotated (Pham et al., 2022). This disease-versus-finding separation is important because it supports explainable or hybrid pipelines in which a system predicts pneumonia while also localizing supporting radiographic evidence.

The annotation quality and cohort structure differ substantially across datasets. PediCXR was manually labeled by pediatric radiologists, but the released labels are effectively single-reader and the paper does not report adjudication or inter-reader agreement (Pham et al., 2022). The Guangzhou/Kermany benchmark is widely used, but several later papers note missing documentation on patient-level splitting, label provenance, or demographic composition (Ridwan, 18 Jul 2025, Hosseinabadi et al., 31 Dec 2025). A plausible implication is that pediatric chest pneumonia systems inherit not only class imbalance but also annotation heterogeneity, dataset-specific shortcuts, and institution-specific bias from their source datasets.

3. Architectural families and system design patterns

One recurrent design choice is the custom CNN trained from scratch. The GAN-augmented pediatric pneumonia classifier uses a custom convolutional architecture with multiple convolutional layers, filters increasing from 32 to 64, max-pooling, batch normalization, and a final dense neuron with sigmoid activation, trained with Adam and binary cross-entropy (Manaf et al., 13 Jul 2025). LightPneumoNet follows the same handcrafted philosophy but is explicitly optimized for parameter efficiency: it uses four convolutional blocks, only 388,082 trainable parameters, and a 1.48 MB footprint, while maintaining high recall on the Kermany benchmark (Chauhan et al., 13 Oct 2025). FA-Net keeps transfer learning but inserts a bespoke attention block—FCSSAM—to suppress irrelevant channels and emphasize discriminative spatial regions before classification (Roy et al., 2024).

A second major family is transfer learning on pretrained CNNs. One comparative study fine-tunes ResNetRS-50, RegNetY-064, and EfficientNetV2-S, with RegNet achieving the best reported performance under a shared pediatric pipeline (Hosseinabadi et al., 31 Dec 2025). The explainable ResNet paper uses ImageNet-pretrained ResNet-50 with a single-neuron binary head and full-network fine-tuning, paired with Grad-CAM and BayesGrad-CAM (Ridwan, 18 Jul 2025). The lightweight weighted-average ensemble model fine-tunes MobileNetV2 and NASNetMobile, then combines their probabilities by a convex weighted average $$0.45\,Y_{\text{MobileNetV2}} + 0.55\,Y_{\text{NASNetMobile}}$$ (Nettur et al., 27 Jan 2025). The federated study chooses a relatively simple ImageNet-pretrained ResNet18 with a custom head, emphasizing that strong pediatric performance can arise from distributed training strategy rather than architectural novelty alone (Jimenez-Gutierrez et al., 12 Nov 2025).

A third family is hybrid and transformer-based modeling. The CNN–ViT fusion paper removes the classification heads of ResNet34 and MaxViT small, concatenates their learned features, and feeds them to a final dense layer with two outputs, achieving 94.87% accuracy, 1.0 sensitivity, and 0.8632 specificity on the pediatric test split (Angara et al., 2024). RepViT-CXR adapts grayscale chest radiographs to a pretrained ViT-Base through channel replication, preserving compatibility with ImageNet-pretrained RGB patch embeddings rather than redesigning the input stem for one-channel data (Ahmed, 10 Sep 2025). The paper’s key claim is that simple replication into three identical channels is sufficient to obtain 98.95% accuracy, 99.34% precision, 99.21% recall, 99.28% F1-score, and 98.73% AUC on the pediatric pneumonia task (Ahmed, 10 Sep 2025).

Not all pediatric chest pneumonia systems are deep CNN classifiers. The functional-regression paper treats each chest radiograph as a two-dimensional function $X_i(h,w)$, derives data-adaptive basis functions from the image covariance structure, and reduces the scalar-on-image model to a generalized linear model on functional principal component scores (Islam, 2020). On the Kermany pediatric dataset, this approach reaches 0.949 out-of-sample accuracy and 0.988 AUC for pneumonia-vs-normal classification, while the harder viral-vs-bacterial problem remains in the 0.69–0.70 accuracy range with AUC around 0.76 (Islam, 2020). This line of work shows that the category “classification system” in pediatrics includes both deep feature learning and statistically parsimonious scalar-on-image models.

4. Data expansion, imbalance handling, and optimization regimes

Class imbalance is a central implementation concern. In the custom CNN-plus-GAN study, the original Guangzhou/Kermany subset used by the authors contains 1,349 Normal images and 3,883 Pneumonia images, so the minority class is explicitly expanded through both geometric augmentation and synthetic generation (Manaf et al., 13 Jul 2025). The paper evaluates four regimes—Original only, Augmented, Generated, and Original + Augmented + Generated—making imbalance correction the organizing principle of the system rather than an incidental regularization step (Manaf et al., 13 Jul 2025).

Geometric augmentation varies across papers, but several patterns recur. The GAN study uses rotation, zoom, shear, and horizontal flip, with rotation up to 40 degrees (Manaf et al., 13 Jul 2025). The transfer-learning comparison applies random horizontal flipping, small rotations ($\pm 10^\circ$), zooming, and brightness adjustments (Hosseinabadi et al., 31 Dec 2025). The federated-learning study uses a richer Albumentations pipeline with HorizontalFlip, Affine transforms with random scaling $\pm 15\%$ and rotation $\pm 30^\circ$, RandomGamma, RandomBrightnessContrast, GaussNoise, Blur or MedianBlur, OpticalDistortion, GridDistortion, ElasticTransform, RandomFog, and CLAHE, all under an overall probability of 0.75 (Jimenez-Gutierrez et al., 12 Nov 2025). “Pneumonia App” adds CLAHE, random resized crop, perspective transform, and rotation specifically to mimic mobile capture distortions (Deng et al., 2024).

Synthetic data generation is less common but highly explicit where used. The GAN-based pediatric classifier trains a generator to produce 148 × 148 synthetic chest X-rays for 40,000 epochs, using the generated images primarily to compensate for the underrepresented Normal class (Manaf et al., 13 Jul 2025). The best performance in that paper comes from the fully combined balanced regime, where Normal is expanded to 5,000 images and Pneumonia to 5,000 (Manaf et al., 13 Jul 2025). The paper does not specify the GAN variant or provide FID/IS-style realism metrics, which makes the augmentation conceptually clear but methodologically underspecified.

Other systems address imbalance through sampling or weighting rather than generation. The ResNet-50 explainability study augments the minority Normal class on disk until it matches the pneumonia count, while also using a weighted random sampler and class-weighted BCEWithLogitsLoss based on the original class frequencies (Ridwan, 18 Jul 2025). LightPneumoNet uses explicit class weights—class_0 (Normal) = 2.0, class_1 (Pneumonia) = 1.2—rather than oversampling (Chauhan et al., 13 Oct 2025). The viral/bacterial study downsamples all classes to 1,500 images per class, producing a balanced 4,500-image working dataset before training and SVM evaluation (Arizmendi et al., 3 Mar 2025). In domain-adaptive adult-to-pediatric transfer, the three-branch contrastive framework enforces a 1:1 ratio between pneumonia and non-pneumonia and a 1:1 ratio between pediatric and adult CXR images at each iteration, using focal loss plus contrastive and embedding-alignment terms to control both class imbalance and domain mismatch (Zunaed et al., 2024).

5. Evaluation regimes, explainability layers, and deployment forms

The most common evaluation metrics are accuracy, precision, recall, and F1-score. One paper explicitly reports

$$\text{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN},\qquad \text{Precision}=\frac{TP}{TP+FP},\qquad \text{Recall}=\frac{TP}{TP+FN}$$

and includes a malformed but clearly intended F1 expression for the harmonic mean of precision and recall (Manaf et al., 13 Jul 2025). Other studies add ROC-AUC, specificity, Cohen’s $\kappa$, MCC, or ECE depending on the task (Ridwan, 18 Jul 2025, Jimenez-Gutierrez et al., 12 Nov 2025, Jayakody et al., 4 May 2026). Reported performance varies sharply with task formulation: binary normal-vs-pneumonia classification is generally strong, whereas bacterial-vs-viral or broader multiclass differentiation is substantially harder.

Several benchmark results illustrate this gradient. The combined original–augmented–GAN CNN reaches 0.86 accuracy and 0.89 F1-score on the binary pediatric task (Manaf et al., 13 Jul 2025). The RegNet-based transfer-learning pipeline reports 92.4 accuracy and 90.1 sensitivity, although that paper also contains a confusion-matrix inconsistency relative to its headline accuracy (Hosseinabadi et al., 31 Dec 2025). The ResNet-50 plus BayesGrad-CAM model reports 95.94% accuracy, 96.76% F1, and 98.91% ROC AUC on the pediatric test set (Ridwan, 18 Jul 2025). LightPneumoNet obtains 94.23% accuracy, 91.94% precision, 99.49% recall, and 95.57% F1 with only 388,082 parameters (Chauhan et al., 13 Oct 2025). RepViT-CXR reports 98.95% accuracy and 99.28% F1-score on the pediatric pneumonia dataset (Ahmed, 10 Sep 2025). By contrast, FA-Net’s more clinically ambitious 3-class task reaches 79.79% accuracy, underscoring the difficulty of subtype separation relative to binary screening (Roy et al., 2024).

Explainability is unevenly developed across the literature. The ResNet-50 study applies Grad-CAM and BayesGrad-CAM to layer4, with BayesGrad-CAM approximating a posterior expectation over model parameters via 20 stochastic forward passes and generating both a mean saliency map and an uncertainty map $U^c$ (Ridwan, 18 Jul 2025). The paper reports that 83% of false positives showed high uncertainty $(U^c > 0.4)$ in critical regions, compared with 12% of true positives, suggesting a triage role for explanation uncertainty (Ridwan, 18 Jul 2025). The older pediatric ensemble paper averages heatmaps across five CNNs and also computes a pixel-wise standard deviation map as an uncertainty measure (Liz et al., 2020). “Pneumonia App” compares Grad-CAM and Score-CAM qualitatively and selects Score-CAM for the mobile interface because it more plausibly highlights suspected lesion regions for the chosen backbone (Deng et al., 2024). At the same time, many systems either omit explainability altogether or do not quantitatively validate heatmaps against expert region annotations.

Deployment spans web, mobile, offline triage, and federated network training. The CNN-plus-GAN system is wrapped in a Flask web application that accepts a JPEG chest X-ray, applies the training-time preprocessing, and returns Normal or Pneumonia together with a probability score; the paper also links both a GitHub repository and a Hugging Face Space (Manaf et al., 13 Jul 2025). “Pneumonia App” exposes a mobile workflow in which an Android client invokes a backend model service through Python Flask, displays the predicted class, and overlays a heatmap on the report page (Deng et al., 2024). MultiSense-Pneumo is designed to run fully offline on standard laptop-class hardware, with a multimodal fusion rule

$$0.45\,Y_{\text{MobileNetV2}} + 0.55\,Y_{\text{NASNetMobile}}$$0

and fixed weights $$0.45\,Y_{\text{MobileNetV2}} + 0.55\,Y_{\text{NASNetMobile}}$$1, $$0.45\,Y_{\text{MobileNetV2}} + 0.55\,Y_{\text{NASNetMobile}}$$2, $$0.45\,Y_{\text{MobileNetV2}} + 0.55\,Y_{\text{NASNetMobile}}$$3, $$0.45\,Y_{\text{MobileNetV2}} + 0.55\,Y_{\text{NASNetMobile}}$$4 (Jayakody et al., 4 May 2026). In contrast, the federated study focuses not on user-facing inference but on privacy-preserving collaborative training across five simulated hospitals using FedAvg, where the federated model reaches 0.900 accuracy, 0.917 F1-score, and 0.966 ROC-AUC under strong non-IID label skew (Jimenez-Gutierrez et al., 12 Nov 2025).

6. Limitations, reproducibility gaps, and emerging directions

A defining characteristic of this literature is that many systems are technically detailed enough to be reproducible in outline but not in exact implementation. The GAN paper omits the GAN variant, generator/discriminator architecture, and synthetic-image quality protocol (Manaf et al., 13 Jul 2025). The RegNet/ResNetRS/EfficientNetV2 comparison leaves unresolved the exact class composition of the 1,000-image subset and contains inconsistencies between its confusion matrix and headline metrics (Hosseinabadi et al., 31 Dec 2025). The ResNet-50 explainability paper does not state how dropout was inserted for Monte Carlo inference, even though BayesGrad-CAM depends on it (Ridwan, 18 Jul 2025). RepViT-CXR reports inconsistent early-stopping patience and does not clearly separate validation from test monitoring (Ahmed, 10 Sep 2025). A plausible implication is that many pediatric pneumonia systems remain closer to research prototypes than to rigorously specified reference implementations.

External validity is another major limitation. Several papers rely on a single public source dataset derived from a single medical center, which constrains generalizability across institutions, scanner vendors, acquisition protocols, and patient populations (Manaf et al., 13 Jul 2025, Ridwan, 18 Jul 2025, Nettur et al., 27 Jan 2025). Even PediCXR, despite its breadth and lesion-level annotations, is effectively a single-country pediatric dataset and does not include adjudication or inter-reader agreement measurements for the released labels (Pham et al., 2022). The federated-learning study demonstrates robustness to synthetically imposed non-IID skew, but all nodes are simulated from one public pediatric dataset rather than from truly independent hospital cohorts (Jimenez-Gutierrez et al., 12 Nov 2025).

Current research directions in the supplied corpus move along three axes. The first is domain adaptation and data expansion, exemplified by adult-to-pediatric three-branch contrastive learning, which improves pediatric test AUROC from 0.8348 under naive joint training to 0.8464 with contrastive and embedding-alignment losses (Zunaed et al., 2024). The second is privacy-preserving collaboration, where federated learning addresses institutional data silos without moving pediatric radiographs (Jimenez-Gutierrez et al., 12 Nov 2025). The third is multimodal clinical assistance, ranging from symptom–cough–speech–image fusion for triage (Jayakody et al., 4 May 2026) to a pediatric pneumonia-centered multimodal LLM that generates reports and records from chest X-rays, CT scans, and text (Tian et al., 2024). These developments suggest that the pediatric chest pneumonia classification system is evolving from a narrow image classifier into a broader pediatric respiratory decision-support stack, but the recurring absence of external validation, calibration analysis, radiologist comparison, patient-level leakage control, and workflow integration evidence indicates that clinical deployment remains a substantially stricter standard than benchmark performance.