Fig. 5: Evaluation of nuclear segmentation performance using various quantitative metrics of pre-trained deep learning models on ROIs of varying nuclear density from the merged dataset.

ROIs that have a nuclear density less than the lower quartile of the merged dataset (A) are regarded as sparse while ROIs that have a nuclear density more than the upper quartile (D) are regarded as dense. ROIs with nuclear density less than the median (B) and more than the median (C) are also analyzed. (i) Mean F1-score (averaged across ROIs) at varying IoU thresholds, with area under the curve shown. A higher IoU threshold results in a stricter condition for classifying a predicted nucleus as a true positive. Mean (averaged across ROIs) with error bars showing 95% confidence interval of (ii) F1-score, (iii) precision, (iv) recall, (v) Jaccard index are evaluated at an IoU threshold of 0.5. The IoU threshold of 0.5 is the most lenient threshold required to ensure a maximum of one true positive predicted nucleus for each ground truth nucleus.