Correcting results to improve your model
Last updated
Last updated
As you acquire different types of images, from different cell lines, procedures, or other variation, you may notice that for some new images, your results may not be 100% accurate. This is often the case when your initial training set didn't capture the variation seen within the images that you are running.
Biodock provides an easy way to take these results, import them as labels into your project, and only correct labels that may be missing or incorrect to improve performance in your model.
You can find the Add to AI Project button right above the image viewer. Click the button to start importing predictions for this option.
A modal will appear. If you've changed your class names, you may need to respecify which classes will be mapped to which labeling classes. In most cases, you'll want to import them back to the same project.
Once you click the blue button in the bottom right, the modal will close, and you'll soon get a notification in platform that the import succeeded. If you go back to the original project, you'll be able to open the imported image:
You should correct one or more of these imported tiles from this image, especially focusing on areas where the model underperformed. Make sure that every object is labeled correctly in the tile. Then you can mark these as fully-labeled by clicking on the tile checkbox.
Finally, once you've imported enough images and corrected more tiles, you should train another version! The new version should perform much better on images that look like the ones you corrected.
