Deep learning for Covid19 Martin B Eriksen Deep
Deep learning for Covid-19 Martin B. Eriksen
Deep learning on CT scans - Prior art - - Deep learning is able to pick up very subtile paterns in CT images. Already known to be useful to detect lung abnormalities: lung tumors, tuberculosis and other pulmonary abnormalities. Literature on brain and lung scans. Source
Motivation for use on Covid-19 - - - Huang 2000 - Infected patents shows signs. X-rays are inexpensive and does not require samples to be distributed analyzed AI tools lowers the need for radiologists, which might also be sick. Companies are working on developing solutions (wired article) Source: The New York times
Prior art - Based on 46046 CT scans from Wuhan, researchers found: “The deep learning model showed a comparable performance with expert radiologist, and greatly improve the efficiency of radiologists in clinical practice. ” - Open repository with some CT scans Exists prior open work on applying DL to CT scans (blog post) Image from the repository
Test appling DL to data. - Running and existing code (fast. ai+Py. Torch). Resnet 34 architecture pretrained on imagenet. Training on GPU is quite fast. Train once, unfreeze some layers and then train more. Right: Some statistic. The error rate is for the test set.
DL results - Relatively high accuracy, 5% errors Right: The confusion matrix. Only 80 particles in the test set. These results might be made more robust with a larger training set.
Discussion - Already exist prior work. Is IFAE entering too late? Spain has a large number of patients. Can local data be considered a resource? We are not doctors or medical researchers. Are we amateurs attempting to help, but creating noise? If we did such a project, what would we gain?
- Slides: 7