Final Year Research Project

AI System for Lung Nodule Follow-Up

Making lung cancer screening more accessible and efficient with cutting-edge AI technology

🎓 Research Team

E/19/366 – W.A.M.P. Senevirathne
e19366@eng.pdn.ac.lk
E/19/224 – M.M.S.H. Madhurasinghe
e19224@eng.pdn.ac.lk
E/18/059 – D.M. De Silva
e18059@eng.pdn.ac.lk

👨‍🏫 Supervisors

Dr. Chathura Weerasinghe
Mr. B.A.K. Dissanayake

📋 Abstract

Lung nodules, potential early signs of lung cancer, are typically detected via CT scans, but their high cost and radiation exposure limit frequent use. Chest X-rays (CXRs) are more accessible but often miss small nodules, leading to delayed diagnoses.

This project develops an AI-powered system for detecting and tracking lung nodules using CXRs, reducing reliance on CT scans. It employs deep learning-based Computer-Aided Detection (CAD) techniques, utilizing Digitally Reconstructed Radiographs (DRRs) to train the AI model for enhanced detection and follow-up.

By integrating feature extraction, image registration, and deep learning, the system offers a reliable, automated solution for lung nodule monitoring. Validated against real-world datasets and radiologist interpretations, the AI system aims to improve early detection and tracking, aiding clinical decisions while minimizing unnecessary CT referrals.

🔬 Related Works

⚠️

Challenge Identified

Early studies highlighted CXR limitations in detecting small nodules

🧠

AI Advancement

Deep learning (CNNs, YOLOv4) significantly advanced automated detection

🏥

Clinical Integration

AI models integrated into hospital PACS for real-time detection

🌍

Global Impact

Making screening accessible in resource-limited settings

⚙️ Methodology

System Overview Diagram

Proposed Workflow

1
Data Collection & Anonymization

Anonymize CT scans with nodules and corresponding CXRs using advanced privacy protection techniques.

2
DRR Generation

Convert CT volumes to Digitally Reconstructed Radiograph (DRR) images that mimic real X-rays.

3
Nodule Projection

Map annotated nodules from CT scans to DRRs for supervised learning and ground truth establishment.

4
AI Model Training

Train deep learning models (CNNs, YOLO) to detect and localize nodules with high accuracy.

5
Follow-up Comparison

Align CXRs over time using image registration to detect changes in nodule size and position.

6
Structured Reporting

Generate millimeter-scale, image-annotated reports for radiologists with clinical insights.

💻 Implementation Details

🔧 Technology Stack

PyTorch TensorFlow SimpleITK MONAI OpenCV Docker FastAPI Python

🖥️ Infrastructure

Cloud Computing: Google Colab Pro/Pro+, AWS EC2 with NVIDIA T4/A100 GPUs for high-performance training and inference.

Local Development: Machines with ≥16GB RAM and NVIDIA GTX 1660+ for development and testing.

📊 Datasets

CT Data: LIDC-IDRI, NLST datasets for comprehensive nodule annotations and ground truth.

CXR Data: CheXpert, JSRT, VinDr-CXR datasets for diverse chest X-ray training data.

Preprocessing: Images standardized to 512x512 pixels with DICOM anonymization using advanced deid techniques.

🎯 Predicted Outcomes

🎯

Enhanced Detection

Significantly reduced false negatives in nodule detection

🔍

Feature Optimization

Advanced feature extraction to assist radiologists

🏥

Clinical Support

Reliable AI system supporting radiologist decision-making

💰

Cost Reduction

Lower follow-up costs through efficient screening

☢️

Safer Screening

Reduced radiation exposure for patients

🌐

Global Access

Improved accessibility in resource-limited settings

📝 Conclusion

Our research develops a groundbreaking AI-powered system for lung nodule detection and follow-up using chest X-rays, significantly enhancing accessibility while reducing reliance on expensive CT scans. By leveraging cutting-edge deep learning techniques and innovative DRR-based training methodologies, our system represents a major advancement in medical AI.

The system improves early detection accuracy, automates complex nodule tracking processes, and provides invaluable support to radiologists in clinical decision-making. This research contributes to making life-saving lung cancer screening more accessible, efficient, and cost-effective worldwide.

📚 Publications

This project has not yet resulted in any publications