Are you confident that your computer vision models will generalize to the variations and batch effects in your inference images?
Have you thoroughly validated your models? What variations should you even be testing for?
Sometimes the distribution change causing degraded performance is obvious, like a different imaging device or geographic location. But it can also be hidden or caused by multiple factors.
Gathering a large, diverse training set may be the ideal solution but is often not feasible – especially for medical images.
What do you do?
Perhaps you have a few techniques in your toolkit like color augmentation or normalization to handle the obvious changes in image color.
But have you ever taken a step back to fully understand the batch effects in your data?
Or what types of distribution shifts you might encounter in the future?
Or searched for spurious correlations that may be biasing your results?
Generalizable models are important for multiple reasons:
- So that they will make correct predictions on previously unseen images
- To ensure they work with images from a variety of scanners, labs, patient populations, geographic regions, vegetation types, etc.
- To get regulatory approval for your product
- So that they are unbiased and provide the most valuable information