Detecting COVID-19 Cases with Deep Learning
0 likes193 views
The document presents an overview of the COVID-Net project, which focuses on using deep learning to detect COVID-19 through chest X-rays, detailing its design choices, hyperparameter tuning, and data augmentation strategies. It emphasizes the importance of balancing sensitivity and false negatives in model evaluation while outlining the real-world and research impacts of this project. Furthermore, it encourages contributions of labeled X-ray data and highlights the utility of SigOpt for tuning experiments.
Detecting COVID-19 Cases with Deep Learning
- 1. Detecting COVID Cases with Deep Learning Tuesday, June 9, 2020
- 2. Introducing our panelists Dr. Alexander Wong Chief Scientist, DarwinAI Canada Research Chair, University of Waterloo Dr. Michael McCourt Head of Research, SigOpt
- 3. SigOpt. Confidential. Agenda Problem: Overview of the problem that led to COVID-Net Design: Trade-offs and choices in COVID-Net development Tuning: Setup and results of the tuning process Impact: Real-world and research impact of this project
- 4. SigOpt. Confidential. Problem Overview of the problem that led to COVID-Net
- 5. Problem: Detecting COVID-19 with chest x-rays Pneumonia COVID-19 Source: COVID-Net preprint paper on Arxiv, https://arxiv.org/pdf/2003.09871.pdf
- 6. SigOpt. Confidential. Design Summary of trade-offs and choices in the model development process
- 7. How it works: Quick introduction to GenSynth Source: DarwinAI, http://www.darwinai.com
- 8. Data Source: COVID-Net preprint paper on Arxiv, https://arxiv.org/pdf/2003.09871.pdf
- 9. Model Source: COVID-Net preprint paper on Arxiv, https://arxiv.org/pdf/2003.09871.pdf
- 10. Results Source: COVID-Net preprint paper on Arxiv, https://arxiv.org/pdf/2003.09871.pdf
- 11. Explainability-Driven Analysis Source: COVID-Net preprint paper on Arxiv, https://arxiv.org/pdf/2003.09871.pdf
- 12. SigOpt. Confidential. Tuning Discussion on the hyperparameter tuning and evaluation process
- 13. How it works: Quick introduction to SigOpt
- 14. Data Augmentation There are only 183 total COVID-19 examples (as of mid-April), but 14000 non-COVID examples. ● To help treat this imbalance, we add a fixed fraction of COVID-19 examples to each batch during stochastic gradient descent. ● Data augmentation is a necessary element to produce more COVID-19 “knowledge” for our network to learn. ○ Zooming ○ Translation ○ Rotation ○ Brightness ○ Trimming Source: SigOpt Blog Post, https://sigopt.com/blog/parametrizing-data-augmentation-in-covid-net-development/
- 15. Tuning COVID-Net Olivia and Mike met with Linda to learn about the design of COVID-Net. ● In particular, we wanted to learn how parameters of the training process and data augmentation were chosen. ● The only work that had been done thus far was a manual search. ○ Linda confirmed that the hyperparameters are important, but did not thoroughly explore the situation. ● We identified 7 parameters to tune in a multimetric experiment. Data augmentation - 4 parameters Loss function - 1 parameter Training process - 2 parameters Source: SigOpt Blog Post, https://sigopt.com/blog/parametrizing-data-augmentation-in-covid-net-development/
- 16. Results Doctors have said that high sensitivity is preferable to high PPV -- this has the effect of minimizing the false negatives. Source: SigOpt Blog Post, https://sigopt.com/blog/parametrizing-data-augmentation-in-covid-net-development/
- 17. Analysis Our real value in this project is being able to produce insights about the relevance of certain parameters to the success of the model building process.
- 18. SigOpt. Confidential. Impact Discussion on the real-world and research impact of this work
- 19. SigOpt. Confidential. Explore the tuning experiment in SigOpt’s dashboard Find the public SigOpt experiment here Use SigOpt at no cost in your academic research sigopt.com/edu Contribute labeled x-ray data to COVID-Net https://figure1.typeform.com/to/lLrHwv Read about other exciting DarwinAI projects DarwinAI Success Stories
- 20. Thank you