A Comparative Study of Deep Learning Architectures for Chest X-Ray Image Classification

Comparative Analysis of Deep Learning Models for Chest X-Ray Image Classification

Author: Akrash Noor, Saba Latif & Hifzun Nisa | Published: December 21, 2025 | Category: AI · Medical Imaging · Deep Learning

Introduction

Medical imaging has become an important aspect in the contemporary healthcare and especially the diagnosis of thoracic diseases utilizing the images of chest X-ray.In recent years, artificial intelligence has advanced significantly, and convolutional neural networks that are based on deep learning have demonstrated impressive results in the domain of automated disease detection and classification. This paper compares and contrasts several deep learning models that are trained on chest X-ray data with PyTorch and TensorFlow in their accuracy, generalization, and computational efficiency.

Deep Learning Models Used

• Custom Convolutional Neural Networks (CNN)
• ResNet (Residual Networks)
• DenseNet
• VGG-style architectures

Methodology

The data is a collection of labeled chest X-ray images of different pulmonary conditions. Preprocessing of images was done on the images to increase the robustness of the models, through use of normalization, resizing and augmentation.

Cross-entropy loss was used to train it and Adam and SGD optimizers were used to optimize it. The metrics used to conduct performance evaluation included accuracy, precision, recall and F1-score.

Framework Comparison

PyTorch and TensorFlow were trained and evaluated using the same architectures. PyTorch was more flexible and easier to debug, whereas TensorFlow was better at deployment and execution optimization.

Results & Findings

Results of the experiment show that more complex architectures like ResNet and DenseNet are more successful in classifying images than shallow CNNs. More basic models in turn have a lower inference time, which means that they can be used in low-resource settings.

Real-World Applications

• Early diagnosis of lung infection and pneumonia.
• Artificial intelligence-aided clinical decision support.
• Distant diagnosis in under-resource areas.

Project Repository

Source code and experiments are available on GitHub:
View Project on GitHub

Conclusion

Deep learning has the potential to revolutionize medical imaging diagnostics. This comparative study highlights the strengths and limitations of various CNN-based models and frameworks, providing valuable insights for researchers and practitioners.

---

Author: Akrash Noor | Saba Latif | Hifzun Nisa
AI Researcher | Machine Learning Engineer | Medical Imaging Enthusiast