Best Practice

In QA's Best Practice, we tell users how to use QA when annotating digital pathology data. This practice includes the preparation of slides, annotation tricks, training tricks, etc.

---

1. What magnification of images should be uploaded?

Theoretically, users could upload the image slides of any magnification. However, we recommend dividing the WSIs of different primitives into different magnifications to get the best segmentation result. 40X is recommended for small scales tissues like cell nuclei, lymphocytes, and glomeruli. 10X is recommended for medium scale tissue like tubules. 10X is recommended for large scale tissues like epithelium and stroma. In general, the larger the tissue scale is, the smaller the magnification is. Improper magnification would confuse the classifier.

Different layers of WSIs have different magnification, depending on the base magnification that the image was scanned at, and the way in which the scanner stores the pyramids.

Please check OpenSlide Python documentation for more details @ https://openslide.org/api/python/#

2. When should a user click the train button?

The classifier needs 2 annotated (1 training patch + 1 testing patch) to start. We recommend annotating 4 patches to start model 1. Then the user could train the nth model when having 2^(n+1) patches in the dataset. We do not recommend training a new model after annotating every single ROI because the model would converge similarly by adding only one more ROI when having many annotations in the data set. That is why we designed our selection rectangle in the form of the 2^n, which corresponds to the logarithm growth in the requirement of annotation data.

Note:

• We assume each patch is a size 256 x 256. Annotating 1 patch of 512 x 512 is equivalent to annotating 4 patches of 256 x 256.
• When the digital slides are of excellent quality, we could change 2 in the 2^(n-1) to a larger integer. In this case, the user would expect the behavior of the classifier converges faster.
• The future version of QA will automate this process.

3. How should the user pick which regions to annotate?

QA provided an embedding plot, helping to select patches for annotation dispersed in the model space. Selecting these patches could help the classifier performs more robustly.

Another way to choose regions is to select the false positive or negative regions of predictions to annotate. (E.g., red circles)

4. How to set numepochs in the training process?

QA employs u-net modeling. The user could set the numepoch in the config.ini file. QA also provides the options for setting num_epochs_earlystop, the early stop epoch number, so the model would stop training when reaching early stop criteria. (e.g., the model is already convergent)