PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comparative Study of GAN-Based Deep Learning Models for Pneumonia Detection in Chest X-Rays

Roshan Paudel, Aashish Ghimire, Pramod Acharya

arXiv:2609.06276v1cs.CV

TL;DR

Limited labeled pneumonia data motivates evaluating synthetic-image augmentation alongside established chest-X-ray classifiers. The study compares a custom CNN with MobileNetV2, VGG19, and ResNet50, finding different accuracy and pneumonia-recall strengths across models. Its VGG19 experiment does not demonstrate a validation benefit from GAN augmentation.

  • Problem

    Limited labeled medical data, particularly positive pneumonia cases, motivates investigating additional ways to supplement training images.

    Source: §I · p. 1

  • Method

    The study compares a custom CNN with MobileNetV2, VGG19, and ResNet50 and tests GAN-generated pneumonia images combined with real training data.

    Source: §I · p. 1

  • Results

    MobileNetV2 achieved 88% accuracy, while the CNN achieved 92.67% pneumonia recall and 79.43% precision; VGG19 GAN augmentation showed no validation-performance improvement.

    Source: §V · p. 6, §VI · p. 8

  • Takeaways & Limitations

    The classifier comparison shows differing accuracy and pneumonia-recall strengths, while the augmentation results support further evaluation before claiming a benefit from synthetic data.

    Source: §V · p. 6, §VI · p. 8

    Larger and more diverse datasets, clearer evaluation-protocol reporting, and external validation are needed before conclusions about clinical applicability.

    Source: §VI · p. 8

Abstract

from arXiv · show

This study evaluates pneumonia classification in chest X-rays using VGG19, MobileNetV2, ResNet50, and a custom CNN, and explores Generative Adversarial Network (GAN)-based synthetic data augmentation. MobileNetV2 achieved the highest reported accuracy of 88% with balanced class-wise performance. The custom CNN achieved pneumonia recall of 92.67% and precision of 79.43%, highlighting a precision-recall trade-off. Accuracy, F1-score, precision, recall, confusion matrices, and training curves were used to assess performance. Synthetic pneumonia images were combined with real images to investigate whether augmentation could improve classification performance. In the reported VGG19 comparison, augmented-data training accuracy reached approximately 100%, while validation accuracy remained near 50%, below the real-data validation accuracy. This experiment therefore did not demonstrate a validation-performance benefit from GAN augmentation. The classifier comparison highlights differences in accuracy and pneumonia recall, while the augmentation experiment indicates the need for further evaluation of synthetic-image quality and training settings.

I. INTRODUCTION

The study addresses limited labeled pneumonia data by comparing CNN classifiers and investigating whether GAN-generated X-rays can supplement real training images. Prior augmentation and imaging studies motivate evaluation, but synthetic-image similarity alone does not establish diagnostic usefulness.

I. INTRODUCTION

  • Limited labeled medical data, especially positive pneumonia cases, motivates additional augmentation strategies beyond rotation and flipping.The introduction links timely pneumonia diagnosis with effective treatment.Source: §I · p. 1
  • GAN-generated X-rays are investigated to add image variation, but image-similarity scores alone cannot establish diagnostic validity or training usefulness.The text says whether synthetic data improves generalization requires experimental assessment.Source: §I · p. 1

II. RELATED WORK

  • Chataut et al. explored shape-aware thoracic edge maps for pulmonary abnormality classification, combining structural and texture information.This related work concerns pulmonary abnormality screening rather than the study’s pneumonia classifier comparison.Source: §I · p. 1

I. INTRODUCTION

  • The study compares a custom CNN with MobileNetV2, VGG19, and ResNet50 while exploring synthetic-data augmentation for pneumonia classification.Performance is assessed with accuracy, precision, and recall.Source: §I · p. 1

II. RELATED WORK

  • Traditional rotation, flipping, and cropping add image variation, though their ability to represent subtle disease features may depend on the dataset and task.Their effectiveness is described as task-dependent.Source: §II · p. 1
  • Prior GAN-based chest-X-ray studies report promising image similarity and increased classification performance, while synthetic-image training remains challenging.The cited discussion cautions that SSIM and PSNR measure image similarity, not equivalent diagnostic quality.Source: §II · p. 1
  • The transfer-learning models use pretrained ImageNet backbones with frozen backbone weights and task-trained output layers; the reported models differ across accuracy, precision, and recall.The described training setup uses Adam at learning rate 0.001, dropout 0.5 in the custom CNN, and 15 epochs.Source: §II · p. 2

III. DATA

The study uses a labeled pediatric chest-X-ray dataset containing Normal and Pneumonia cases, then supplements pneumonia training data with GAN-generated images. Its classification reports cover 300 evaluated images split evenly between the two classes.

III. DATA

  • The Kaggle Chest X-ray Dataset contains 5,863 labeled Normal or Pneumonia images focused on children aged 1–5.Images were collected at Guangzhou Women and Children’s Medical Center, with diagnostic labels reviewed by two physicians and a third reviewer.Source: §III · pp. 2–3
  • GAN-generated images supplement available pneumonia images, while SSIM and PSNR are discussed without numerical results establishing synthetic-image quality.The classification reports cover 300 evaluated images, with 150 per class.Source: §III · p. 3
  • The example figure presents chest X-rays from the pneumonia and normal categories.Source: Figure 5 · p. 3

IV. METHODOLOGY

The methodology combines real and GAN-generated pneumonia images, then trains custom and transfer-learning classifiers for binary chest-X-ray classification. Images are standardized for processing, and performance is assessed with classification metrics and confusion matrices.

IV. METHODOLOGY

  • The pipeline generates synthetic pneumonia X-rays, merges them with real images, and trains classifiers to investigate whether synthetic images can supplement training data.The evaluation uses accuracy, precision, recall, F1-score, and confusion matrices.Source: §IV · p. 3

A. Data Preprocessing

  • Real chest X-rays are resized to 512 × 512 pixels and normalized before GAN training and classification.The real dataset is divided into training, validation, and test sets.Source: A. Data Preprocessing · p. 3

B. Classification Models

  • The classifier set includes a custom CNN and MobileNet, VGG19, and ResNet50 architectures.The text describes CNN convolutional, pooling, and fully connected layers; MobileNet depthwise separable convolutions; VGG19 repeated 3×3 convolutions; and ResNet50 residual connections.Source: B. Classification Models · pp. 3–4

C. Synthetic Data Generation Using GANs

The GAN was trained on pneumonia-labeled chest X-rays to generate synthetic images, which were combined with real images for an augmented dataset. The reported sample contains artifacts and blurred detail, and the outputs do not establish diagnostic feature preservation.

C. Synthetic Data Generation Using GANs

  • The GAN used a generator to produce pneumonia-positive X-rays resembling training images and a discriminator to distinguish real from synthetic images.The two networks were trained adversarially; the generator aimed to approximate image intensity, texture, and structure.Source: C. Synthetic Data Generation Using GANs · p. 4
  • The shown synthetic sample contains visible artifacts and blurred detail, while the reported outputs do not establish preservation of diagnostically relevant pneumonia features.SSIM and PSNR are described as similarity measures, but numerical results and the reference-image pairing procedure are not provided.Source: C. Synthetic Data Generation Using GANs · p. 4, Figure 6 · p. 4
  • Synthetic pneumonia images were combined with the original real dataset to create an augmented dataset for classifier training.The purpose was to address data sparsity and test whether additional images could improve classifier performance.Source: C. Synthetic Data Generation Using GANs · p. 4

D. Model Training and Transfer Learning

The transfer-learning models used pretrained feature backbones with a trained classification output layer, and the paper reports selected hyperparameter settings. A separate hyperparameter-search procedure is not documented.

D. Model Training and Transfer Learning

E. Advantages of Selected Hyperparameters and Potential Improvements

  • Adam and a learning rate of 0.001 were reported, but the results do not establish that this learning rate was optimal.The custom CNN used dropout of 0.5; the pretrained architectures used GlobalAveragePooling2D, while the CNN used MaxPooling2D.Source: E. Advantages of Selected… · p. 5

F. Potential Hyperparameter Tuning

The section lists possible hyperparameter variations for future experiments, including optimizer, batch size, learning schedule, regularization, and model structure. Experiments are needed to determine whether they improve accuracy, robustness, or generalization.

F. Potential Hyperparameter Tuning

  • Dynamic learning-rate schedules and He or Xavier initialization are proposed as alternatives to investigate convergence behavior.The suggested schedules include cosine annealing and step decay.Source: F. Potential Hyperparameter Tuning · pp. 5–6
  • L2 regularization, different dropout rates, and conventional image augmentation are proposed as ways to explore overfitting and input variability.The augmentation examples are rotation, flipping, and cropping, particularly for MobileNetV2 or ResNet50.Source: F. Potential Hyperparameter Tuning · pp. 5–6
  • The section also proposes varying epochs, using validation-based early stopping, testing focal loss, and tuning CNN layer width or depth.Experiments would be needed to determine whether these changes improve accuracy, robustness, or generalization to unseen data.Source: F. Potential Hyperparameter Tuning · pp. 5–6
  • Testing RMSProp or momentum SGD and varying batch size could explore effects on performance and gradient estimation.The passage notes that smaller batches can introduce regularization effects, while larger batches stabilize training.Source: F. Potential Hyperparameter Tuning · pp. 5–6

G. Evaluation Metrics

The evaluation uses accuracy, precision, recall, and F1-score to assess classification performance. SSIM is identified as an image-similarity measure, but numerical SSIM results are not reported.

G. Evaluation Metrics

  • Accuracy is the proportion of correct predictions, while precision is the proportion of true positives among all positive predictions.Source: G. Evaluation Metrics · p. 6
  • Recall is the proportion of true positives among actual positive cases, and F1-score is the harmonic mean of precision and recall.Source: G. Evaluation Metrics · p. 6
  • SSIM measures structural similarity between images, but numerical results are not reported and similarity alone does not establish diagnostic realism.Source: G. Evaluation Metrics · p. 6

V. RESULTS AND DISCUSSION

The section compares four classifiers and reports different leaders for accuracy and pneumonia recall; the VGG19 real-versus-augmented comparison shows no validation-performance gain.

V. RESULTS AND DISCUSSION

A. Experimental Analysis

  • VGG19 augmented-data training accuracy reached approximately 100%, while validation stayed near 50%, below real-data validation at approximately 75–79%.The displayed experiment does not show a validation-performance improvement from GAN augmentation.Source: A. Experimental Analysis · pp. 6–7

V. RESULTS AND DISCUSSION

  • CNN missed 11 pneumonia cases and classified 36 normal cases as pneumonia, compared with 19 and 17, respectively, for MobileNetV2.Source: §V · p. 6, Figure 10 · p. 7

A. Experimental Analysis

1) Interpretation of ROC Curve and Precision-Recall Curve:

The section explains ROC and precision-recall curves as complementary threshold-based views, contrasting class separation with positive-class precision and recall.

1) Interpretation of ROC Curve and Precision-Recall Curve:

VI. CONCLUSION

MobileNetV2 led reported accuracy, while the custom CNN led pneumonia recall; the reported GAN-augmentation comparison did not demonstrate improved generalization.

VI. CONCLUSION

  • The VGG19 comparison showed lower validation accuracy with GAN-augmented data than with real data, so the results do not demonstrate improved generalization.Source: §VI · p. 8
  • Larger, more diverse datasets, clearer evaluation-protocol reporting, and external validation are needed before drawing clinical-applicability conclusions.Source: §VI · p. 8
  • The custom CNN achieved 92.67% pneumonia recall, while MobileNetV2 achieved 88% accuracy with balanced class-wise metrics.Source: §VI · p. 8
  • Extending GAN training to 500 epochs remains a proposed experiment, not an established route to improved image quality or pneumonia recall.Source: §VI · p. 8

VII. FUTURE WORK

Future work proposes refining GAN image quality and training settings, then testing classification effects with matched comparisons and broader evaluation.

VII. FUTURE WORK

  • Future work proposes conditional GANs or auxiliary classifier GANs to reduce visible artifacts and evaluate whether their outputs improve classification.Source: §VII · p. 8
  • Learning rates, batch sizes, training duration, and regularization could be varied across VGG19, MobileNetV2, ResNet50, and the custom CNN.Source: §VII · p. 8
  • Matched real-data and augmented-data comparisons on larger, more diverse datasets are proposed to test generalization and clinical applicability.Source: §VII · p. 8
  • Possible extensions include lesion localization and image-grounded question answering, with each capability evaluated separately.The cited dental-radiograph work found recognition performance did not translate into comparable localization and counting performance [14].Source: §VII · p. 8
  • If an extension uses an LLM-powered browser or assistant to process external reports, evaluation should address prompt injection.Source: §VII · p. 8
Loading 2609.06276v1…