Covid-19 CT Scan Reader

• 

  original image

• 

  segmentated image

• 

  We are hoping that our model doesn't crash again after crashing 3 times in a row.

Inspiration

Covid-19 doesn’t need an introduction. We all know how much it interfered with not only our daily lives, but our health, and the health of our loved ones.

From early on in the pandemic, radiologists had an inclination that CT scans of patients’ lungs held signs of Covid-19 infection. After all, lungs in these patients had similar damage to those in patients suffering from SARS-CoV-1 and MERS-CoV, other notable coronaviruses. Key identifiers of such damage in CT scans were classified, leading directly into our project. We trained a TensorFlow model for the detection of acute and organizing diffuse alveolar damage in publicly available CT scans. The model operates through image segmentation.

We hope models like this can be used to address multiple diseases (such as the aforementioned coronaviruses). If we can free up the valuable time of trained radiologists and physicians as they address the critical health care needs of the general public, we believe everyone will be better off. Of course, apps like these bring up interesting ethical issues, especially regarding misdiagnosis and health privacy. While we can’t address all of these issues here in this documentation, we thought it prudent to mention these issues and mention that they are of key interest to the members of our team.

What it does

It takes an image from an user and then runs an Unet on the image and outputs the results in an UI

How we built it

We used google colab and python machine learning libraries to do the training of the model. Tkinter is used for the app

Challenges we ran into

We found it difficult to understand the research paper about Unets and learning about UI design was challenging

Accomplishments that we're proud of

We got an Iou score of 90 + percent which was great and the UI looks good also

What we learned

we learned about Unet as well as UI design

What's next for Covid-19 CT Scan Reader

We plan to explore attention methods for improving our Unet performance as well as exploring different backbone structures

Built With

• colab
• google
• python
• sklearn
• tensorflow
• tkinter