Pneumonia Detection: CNN vs. Medical Professionals

This repository contains the code and analyses for the project in Introduction to Intelligent Systems (02461) at DTU, January 2025.

Authors:

• Valdemar Stamm Kristensen (s244742)
• Frederik Lysholm Jønsson (s245362)
• William Hoffmann Hyldig (s245176)

Study line: Artificial Intelligence and Data

---

📌 Project Overview

This project explores the use of Convolutional Neural Networks (CNNs) to detect pneumonia in chest X-ray images and compares the model’s performance to that of two medical doctors.

The motivation lies in pneumonia’s status as a major global health challenge, particularly among children, where fast and reliable diagnosis is crucial. CNNs offer a way to assist doctors in handling the large volume of X-ray data in clinical settings.

---

🧠 Key Methods

• Dataset: Kaggle chest X-ray dataset with 5,116 images (73% pneumonia, 27% normal).
• Preprocessing: Resizing (244×244), normalization, and tensor conversion.
• Model Architecture: Custom CNN inspired by VGG-16, reduced in complexity to prevent overfitting.
• Training Distributions: Compared original (73/27) vs. balanced (50/50) data splits.
• Evaluation: Compared CNN predictions against two non-specialist medical doctors reviewing 100 X-rays each.

---

📊 Results in Brief

• CNN Performance: Achieved 96.08% ± 1.57 test accuracy.
• Medical Doctors: Achieved 67–72% ± ~9 accuracy.
• Observations:
  • 73/27 model achieved highest overall accuracy but showed bias toward pneumonia.
  • 50/50 model was more stable across distributions but required discarding data.
  • Overfitting was observed after epoch 7 → optimal stopping point identified.

These results confirm that CNNs can significantly outperform non-specialist medical professionals in classifying pneumonia from X-rays.

---

📂 Repository Structure

• CNN-model.ipynb → Full Jupyter Notebook with preprocessing, model design, training, evaluation, and visualizations.
• CNN-model.py → Same code as CNN-model.ipynb but in a python file.

---

🔬 Reflections and Future Work

• Dataset limitations: Current data originates from one hospital → limited generalizability.
• Bias risks: Model may perform differently across age, gender, and imaging conditions.
• Future improvements:
  • Larger and more diverse datasets across multiple hospitals.
  • Use of data augmentation (flipping, rotation, contrast adjustments).
  • Certainty scores and explainable AI (XAI) to improve clinical trust and usability.

---

📖 References

• Kaggle Pneumonia Dataset: Mooney, P. (2018). Chest X-Ray Images (Pneumonia).
• Sharma, A. (2024). Pneumonia Detection using VGG16 Transfer Learning.
• UNICEF (2021). Pneumonia statistics on child mortality.
• Additional references listed in the project report.