Fig. 5.

Qualitative performance of NuCLS model on testing sets. The displayed ground truth comes from the pathologist-corrected single-rater dataset. The images are representative of a number of different hospitals in each of the testing sets. Detection and classification performance closely matches the ground truth, and discrepancies are marked by arrows. Not all discrepancies are algorithmic errors, including: (i) adjacent nuclei that could conceivably be viewed as a single nucleus; (ii) missing annotations and (iii) morphologically ambiguous nuclei