- Hypothesis testing: our approach was to develop the models by posing hypotheses and repeatedly testing them at different points in development (e.g., whether domain-inspired data augmentation would improve model performance)
- Iterative approach: we developed several iterations of models using data that was partially annotated using previous iterations and corrected by hand
- Domain knowledge: we liaised with the client to inform them of our data augmentation strategy and guide the criteria for model acceptance
- Define milestones: we defined several important milestones which guided the model development and performance, including: over 95% average dice score in all cases, over 95% average dice on severe COVID cases, submit abstract based on work to a conference, and host model internally.
- Our model performs with >95% average dice over all test cases, compared to benchmark models, for example, the state-of-the-art work by Hofmanninger, which achieved dice scores of around 90% on our data
- Our model can be used to volumetrically distinguish inhalation and exhalation CT scans with >95% F1 score. This allowed other research teams to sort their data automatically, saving days or weeks of tedious manual work
- Our paper was accepted at an international biomedical conference to discuss the benefits of the data augmentation approach we used. This gave us a chance to showcase our expertise and data findings
- Our model is hosted internally, providing lung segmentation to the organization, expediting the use of this model by internal company research teams to accelerate their research.