But is this data useful, can a neural network learn to detect COVID-19 from US images? To answer this question, we build a very simple proof-of-concept ML model, called POCOVID-Net. It consists of a pre-trained VGG net where only the last convolutional layer is refined during training, and a fully connected layer of 64 neurons. The model is trained to distinguish between COVID-19, pneumonia and healthy patients. Augmentation such as rotations, shifts and flips help to prevent overfitting on the dataset that is still rather small for a model with two million trainable parameters.

The preliminary results are very promising — POCOVID-Net achieves an accuracy of 89% to classify into COVID-19, pneumonia and healthy patients. The sensitivity, which is most important in the current situation, is even at 96%. Below you can see the confusion matrices, with absolute values and then normalized along each axis respectively.

#deep-learning #ultrasound #covid-19-testing #covid19 #ai