This project focuses on the detection of pneumonia from chest X-ray images using two different neural network models: a Convolutional Neural Network (CNN) and a Multilayer Perceptron (MLP). These models aim to improve the accuracy and efficiency of pneumonia detection, particularly in settings with limited access to expert radiologists.
Citation: OpenAI. (2023). AI-generated image of pneumonia detection using chest X-rays with neural network. OpenAI's DALL-E.
- Pratheeksha Nath Narikkadan
- Sahithi Sallaram
- Sruthi Atluri
- Supriya Dhamapurkar
The project introduces two neural network models to diagnose pneumonia from chest X-ray pictures. Our CNN model, which is custom-built for this purpose, outperformed the MLP model with an accuracy of 91.02% compared to 76.40% for MLP. We have also developed a user interface that accepts a chest X-ray and predicts the occurrence of pneumonia, including the percentage of congestion. This work not only contributes to the medical field by providing a tool for early detection but also serves as a significant step in the application of machine learning for healthcare improvements.
- CNN (Convolutional Neural Network): This model was built from scratch, focusing on pneumonia detection from chest X-ray images. It achieved higher accuracy and is more efficient in handling image data.
- MLP (Multilayer Perceptron): A classic neural network model that, while slightly less accurate than the CNN, still demonstrates the potential for machine learning in medical imaging.
Details on how to use these models, including the user interface for pneumonia detection, can be found in the respective directories of this repository.
The models were trained and tested on a dataset consisting of chest X-ray images, labeled as either 'Pneumonia' or 'Normal', sourced from Kaggle. Find the Dataset here
For more details on our methodology, findings, and contribution to the field, please refer to our project report included in this repository.
For any queries regarding this project, feel free to reach out to us via email: