Classification and repeatability metrics on a held-aside test set of 8,734 images from the “SEED” dataset.

<p>Our model performs consistently well on “internal” data, even when retrained with added “external” images from the “EXT” dataset. Metrics are reported for each of the model runs involving incremental additions of images from the “EXT” (J8) dataset at the woman level, in a 2n normal (N): 2n indeterminate (I): 1n precancer+ (P) ratio of ground truth class, where n = # of precancer+ women added, as shown on the leftmost column. % values are rounded to 1 decimal place, while numeric values are rounded to 2 decimal places.</p> <p>(DOCX)</p>