Chest X-Ray Classification with MobileNetV2 (Transfer Learning)

This project applies deep learning (CNN transfer learning) to classify chest X-ray images into COVID-19, Normal, and Viral Pneumonia. The goal is to explore the feasibility of an automated clinical triage support approach using radiographic patterns and to analyze clinically relevant errors via sensitivity/specificity and confusion matrices.

Dataset

COVID-19 Chest X-Ray Database (Qatar University & University of Dhaka collaborators). Multisource public CXR collection including COVID-19, Normal and Viral Pneumonia images. License: CC BY 4.0

Objective

To build a baseline deep learning model for multiclass chest X-ray classification that could serve as a triage support tool, and to assess performance using clinically meaningful metrics (confusion matrix, sensitivity/specificity per class, ROC-AUC).

Methods

Modeling approach

Multiclass image classification using Convolutional Neural Networks (CNN) with transfer learning. Pre-trained MobileNetV2 architecture (ImageNet weights) used as a feature extractor. Final classification head trained for three classes: COVID-19, Normal, Viral Pneumonia.

Image preprocessing

Original chest X-ray images are provided at 256×256 pixels resolution. All images were resized to 224×224 pixels to match the canonical input size of the pre-trained MobileNetV2 architecture and ensure efficient transfer learning. Pixel values were normalized to the [0,1]range.

Data splitting

Stratified 80/20 train–validation split performed per class, preserving class proportions.

Training strategy

The convolutional backbone was initially frozen to preserve learned low-level visual features. Only the classification head was trained. Optimization performed using Adam optimizer with early stopping based on validation loss.

Data augmentation

Light, anatomy-preserving augmentations were applied during training:

• Small rotations
• Minor translations and zoom No aggressive geometric or contrast transformations were applied to prevent the model from learning non-clinical artifacts.

Evaluation metrics

• Overall accuracy
• Class-wise precision, recall (sensitivity) and F1-score
• Confusion matrix
• Multiclass ROC-AUC (One-vs-Rest) with macro and weighted averaging

Key Results

• Overall validation accuracy: ~0.90
• Macro ROC-AUC (OvR): ~0.98, indicating strong overall discriminative ability.
• High specificity for pathological classes, particularly COVID-19.

Class-wise performance highlights

• COVID-19 : Specificity: 0.99, Sensitivity (recall): 0.67
• Normal : High sensitivity, slightly reduced precision
• Viral Pneumonia : Balanced sensitivity and specificity

Clinical Interpretation and Triage-Oriented Perspective

The model demonstrates very high specificity for COVID-19 (≈0.99). This indicates that when the model predicts COVID-19, it does so with high confidence and produces very few false positive COVID classifications. In practical terms, the model is highly reliable at recognizing pathological COVID-related patterns when they are present. Conversely, COVID-19 sensitivity is moderate (≈0.67). Analysis of the confusion matrix shows that a subset of true COVID-19 cases are misclassified as Normal. These represent false negative COVID cases, not false positives. Importantly, this behavior is clinically informative rather than purely erroneous.

Triage implications

From a triage perspective, the model’s behavior suggests the following workflow interpretation:

• Cases predicted as COVID-19 can be considered high-confidence pathological studies, suitable for fast-tracking further evaluation, isolation protocols, or confirmatory testing.
• Cases predicted as Normal, particularly in high-prevalence or high-risk contexts, should not be automatically dismissed. The observed false negatives indicate that this group may benefit from secondary review, clinical correlation, or follow-up imaging.

Thus, the model effectively facilitates prioritization of clinical processes, rather than replacing medical judgment.

Clinical trade-off and future optimization

The model is intentionally optimized for high specificity in COVID-19 detection, minimizing false positive diagnoses. This conservative behavior is well suited for clinical decision-support and triage workflows, where avoiding unnecessary escalation of care and resource utilization is critical. Future work will focus on increasing COVID-19 sensitivity through cost-sensitive learning and targeted fine-tuning of deeper convolutional layers, while preserving the model’s strong capacity to identify pathological lung patterns across the entire field of view.

Project Structure

.
├── data/
│   └── raw/
│       └── README_dataset.md       # Original dataset description and references
│
├── models/
│   └── xray_cnn_mobilenetv2.keras  # Trained MobileNetV2-based CNN model
│
├── results/
│   ├── classification_report.txt  # Precision, recall, F1-score per class
│   ├── roc_auc.txt                # Macro and weighted ROC-AUC scores
│   ├── confusion_matrix.png       # Confusion matrix (validation set)
│   ├── accuracy_curve.png         # Training vs validation accuracy
│   ├── loss_curve.png             # Training vs validation loss
│   └── recall_per_class.png       # Sensitivity (recall) per class
│
├── src/
|   └── train_mobilenetv2.py       # End-to-end training and evaluation script
├── .gitignore                     # Git ignore rules (e.g. virtual environment)
└── README.md                      # Project documentation

Note:

Raw image data are not included in this repository. The folder structure is preserved to ensure reproducibility, while images must be obtained from the original public dataset (see Dataset section).

Technical Stack

• Programming language: Python
• Deep Learning: Convolutional Neural Networks (CNN)
• Architecture: MobileNetV2 (transfer learning, ImageNet pretrained)
• Framework: TensorFlow / Keras
• Data pipeline: tf.data API
• Evaluation metrics:
  • Accuracy
  • Precision, Recall (Sensitivity), F1-score
  • Specificity (derived from the confusion matrix)
  • ROC-AUC (One-vs-Rest, macro and weighted)
• Visualization: Matplotlib
• Environment: Linux (Ubuntu / WSL)

How to run

1. Clone the repository
2. Create and activate a Python virtual environment
3. Install dependencies
4. Run the training script

Why this project matters

This project illustrates how deep learning models can be applied to real-world medical imaging workflows, emphasizing not only predictive performance but also clinically meaningful trade-offs such as sensitivity, specificity, and triage prioritization. The focus on error analysis and clinical interpretation reflects the requirements of healthcare AI systems beyond purely technical accuracy.