This project developed a convolutional neural network (CNN) model that identifies pneumonia cases from chest x-rays.
The CNN model achieved a 91.67% accuracy, with a 95% recall on identifying positive cases (95% sensitivity), and a 86% on negative cases (86% specificity) on 120 testing x-ray images.
Examples:
(more examples in /images)
Pneumonia is an infection of the lungs. Its seriousness can range from mild to life-threatening, which is why identifying positive pneumonia cases from the results of the medical tests is important.
| normal | pnemonia |
|---|---|
 |
 |
| Accuracy | Loss | Sensitivity (True positive rate) | False negative rate | False positive rate | Specificity (true negative rate) |
|---|---|---|---|---|---|
| 0.9167 | 0.22 | 0.95 | 0.05 | 0.08 | 0.86 |
-
95% sensitivity in medical tests is genearlly considered a high sensitivity. The model also has a 86% specificity. Therefore, this trained model is good at identifying positive cases correctly, but is more prone to falsely predicting a negative result as positive. However, in the context of pneumonia, this result should be not bad, if not pretty good.
-
86% specificity - although there is room for improvement, this number of specificity is not very low in medical tests
-
95% sensitiivty - a very good sensitivity in real world medical tests
In the case of pneumonia, sensitivity is more important than specifity because pneumonia could be deadly. If we have to choose, we would rather have false poisitives than missing a true positive case.
| model architecture |
|---|
 |
Side note: The same CNN model yielded good results with two different random states. However, one of them turned out to have a 100% sensitivity, which would never happen in real world, and it achieved so by sacrificing specificity (61%). The one kept in this repo is therefore the other one as detailed above.
Credits to this Kaggle dataset for all x-ray images used in this project.