Covid-19 Detection from Chest X-Ray (Tool)
Semplice Tool autoprodotto che, utilizzando una Rete Neurale Convoluzionale (composta da circa 5 milioni di parametri) addestrata con un dataset composto da 206 radiografie di pazienti affetti da Covid-19 e 206 radiografie di persone sane, effettua la classificazione di nuove radiografie al fine di verificare la presenza di Covid.
Disclaimer
This Tool is just a DEMO about Artificial Neural Networks so there is no clinical value in its diagnosis and the author is not a Doctor! Please don’t take the diagnosis outcome seriously and NEVER consider it valid!!!
Info
This Tool gets inspiration from the following works:
- Detecting COVID-19 in X-ray images with Keras, TensorFlow, and Deep Learning
- Fighting Corona Virus with Artificial Intelligence & Deep Learning
- Deep Learning per la Diagnosi del COVID-19
We used 206 Posterior-Anterior (PA) X-Ray images of patients infected by Covid-19 and 206 Posterior-Anterior X-Ray images of healthy people to train a Convolutional Neural Network (made by about 5 million trainable parameters) in order to make a classification of pictures referring to infected and not-infected people. Since dataset was quite small, some data augmentation techniques have been applied (rotation and brightness range). The result was quite good since we got 94.5% accuracy on the training set and 89.3% accuracy on the test set. Afterwards the model was tested using a new dataset of patients infected by pneumonia and in this case the performance was very good, only 2 cases in 206 were wrongly recognized. Last test was performed with 8 SARS X-Ray PA files, all these images have been classified as Covid-19. Unfortunately in our test we got 5 cases of ‘False Negative’, patients classified as healthy that actually are infected by Covid-19. It’s very easy to understand that these cases can be a huge issue.
We are aware the model is suffering of some limitations:
- small dataset (a bigger dataset for sure will help in improving performance)
- images coming only from the PA position
- a fine tuning activity is strongly suggested
Anybody has interest in this project can drop me an email (rosario.moscato@outlook.com) and I’ll be very happy to reply and help.