About Our Project

Welcome to our Chest X-ray AI project, an innovative second opinion application developed by students at McMaster University as part of their capstone project. Our mission is to leverage cutting-edge artificial intelligence technology to assist in the interpretation of chest X-rays, providing a valuable tool for our users.

Our Approach

We have developed a sophisticated AI model using Convolutional Neural Networks (CNNs), specifically employing the DenseNet201 architecture. This model has been fine-tuned and optimized to achieve high performance in identifying various chest conditions from X-ray images.

Our AI was trained on the Stanford CheXpert dataset, which comprises a total of 224,316 chest radiographs of 65,240 patients. For more details on the dataset, please visit the Stanford CheXpert website

Performance Metrics

Our model was rigorously tested on over 33,000 images, demonstrating strong performance with high values in key diagnostic metrics:

Additional Features

Instead of binary "disease present" or "disease absent" outputs, our AI provides probability-based predictions, displayed as percentage likelihoods using progress bars. This allows users to assess the model's confidence in its diagnosis rather than relying on a simple yes/no result.

To enhance clinical interpretation, our application features an advanced visualization tool: a Grad-CAM-based heatmap. This heatmap visually highlights regions of the X-ray that were most influential in the AI's prediction. This visual guidance helps you focus on the most important areas of the X-ray, making it easier to spot potential issues or confirm your observations. This visual guidance can be particularly valuable in identifying subtle abnormalities or confirming the location of suspected pathologies, thereby supporting more informed diagnostic decisions.

Disclaimer

While our AI model has been developed with cutting-edge techniques and shows promising results, it is important to remember that AI is not infallible. Its outputs are intended to support and not replace the clinical judgment of healthcare professionals. We strongly advise consulting a qualified radiologist or medical doctor before making any clinical decisions based solely on our tool.

Open Source

We believe in transparency and open science. Our project's code is available for exploration on our Github Repository. We encourage interested individuals to review our work and gain insights into our development process.

By combining advanced AI techniques with a commitment to responsible and transparent development, we aim to contribute to the field of medical imaging and support healthcare professionals in their critical work. Our open-source approach allows for scrutiny and understanding of our methods, fostering trust and advancement in AI-assisted medical imaging.