CheXNet an algorithm able to compete with radiologists
Xray4ll is a web application developed by Stanford students that integrates a CheXNet algorithm to interpret chest X-rays. This algorithm is the first to simultaneously evaluate chest x-rays for a multitude of possible diseases and to give results consistent with radiologists diagnosis.
Scientists have trained the algorithm to detect 14 different diseases: for 10 diseases, the algorithm performed as well as the radiologists; for three diseases, it gave inferior results to those of the radiologists; and for one disease, the algorithm outperformed the experts.
All you have to do is upload a chest X-ray from your computer or cell phone and wait for the analysis of Xray4ll which can validate or not the presence of the pathology but also visually locate the area of attention highlighted by the algorithm thanks to a visual mapping. This visual mapping is obtained thanks to a second GRAD-CAM algorithm.
CheXNet applications for COVID-19
Researchers have tried to use the CheXNet model to detect COVID-19. The number one priority was to build a dataset with Chest x-rays of patients with COVID-19.
The researchers shared open source databases of chest x-rays on Github to further the research. The goal of the project was to create a public dataset of chest x-rays of patients positive or suspected of having COVID-19 or other viral and bacterial pneumonias (MERS, SARS and ARDS).
🛑 Note: please do not claim diagnostic performance of a model without a clinical study! This is not a kaggle…
The researchers then used the transfer learning technique by taking the pre-trained parameters from CheXNet.
“Transfer learning is an approach to deep learning (and machine learning) in which knowledge is transferred from one model to another.”
Typically, the first few layers of the network (especially for CNNs) are frozen. This allows researchers to perform a full training on the existing model and change the parameters in the very last layers.
“It has been observed that the early layers of a network capture more generic features while the later layers are very specific to a data set.”
The proposed system is able to perform on binary and multi-classification tasks with an accuracy of 98.08% and 87.02%, respectively.
Pie & AI: Real-world AI Applications in Medicine (from 1h23, demo Xray4ll by Bora Uyumazturk)
Transfer Learning par Andrew Ng