TL; DR

I have tried to do deep learning on COVID data using Eisen. It seems that all the pieces needed to experiment with most of the publicly available datasets are already included in the package.

I have coded up a small example showcasing segmentation of lung data from an open dataset of 100 images obtained from the internet. Eisen had already pre-built datasets capable of working with such data and allowed me to train my first model after a couple of minutes.

Event though I don’t think I can produce a state of the art model in less than 5 minutes, I think that having fast prototyping and development tools during this pandemic will help iterate faster. Being able to leverage a community of developers, such as the one around Eisen, who care about building the best possible tools for medical DL, is the key to reach tangible results faster.

---

In the last few weeks we have seen quite a few studies showing what can be done when data about the Covid-19 global pandemic is analyzed using ML/DL methods. A few different public initiatives have been announced and some data has been released.

Although we are still awaiting to have a large dataset of medical images that can be used for crucial tasks such as survival prediction, ICU workload, complications and long term outcome, a few smaller efforts have produced publicly available datasets such as those obtainable here and here.

I have seen quite a lot of buzz about these (and other) datasets, and a few people have started their own GitHub projects to run their models on this data. Unfortunately they have often gone to the lengths of coding everything from scratch which has caused bugs and lack of reproducibility of results.

Since a few weeks, Eisen natively supports two of the most important Covid-19 datasets out there. In order to create a deep learning model trained on this dataset, one needs to write only a couple lines of code and, voilá, a fully functioning, bug-free and reproducible training operation can be started.

I have decided to try out the functionality offered by Eisen in relation to Covid-19 data. There are two datasets being available as of now. They can be instantiated using the following lines for the UCSD dataset:

from eisen.datasets import UCSDCovid19

dataset = UCSDCovid19(
    '/abs/path/to/data',
    'positive',
    'negative',
    transform,
)
# docs here http://docs.eisen.ai/eisen/api.html#eisen.datasets.UCSDCovid19

And the following snippet for the MedSegCovid19 dataset.

#computer-vision #deep-learning #covid19 #medical-imaging #ai #deep learning