Fig. 7: Evaluation of nuclear segmentation by the pre-trained Cellpose, Mesmer, and StarDist models on representative ROIs.

A Breast tissue type and B tonsil tissue type with high nuclear density. (i) Spectrally unmixed mIF data has the (ii) DAPI channel extracted and (iii) ground truth nuclei annotated. The pre-trained models are applied to the DAPI channels to yield (iv) binary nuclear masks, which are (v) overlayed on the ground truth mask for comparison to visualize the model-derived differences in nuclear segmentation. The differences between the models in the (vi) recall and (vii) precision, which both constitute the F1-score are also visualized. All F1-score, precision, and recall metrics are evaluated at an IoU threshold of 0.5, and nuclei that have IoU lower than the threshold are included as false positives in the precision and false negatives in the recall calculation. Scalebar in yellow, 25 µm.