Fig. 5: Expert survey results and the exclusion of regions with severe artifacts for the subsequent H&E Cryo dataset analysis.
From: Deep learning-based image analysis in muscle histopathology using photo-realistic synthetic data
Diverging stacked bar chart showing the quantitative results of an expert survey evaluating the segmentation performances of U-Netreal and U-Netsynth on the unlabeled H&E Cryo dataset. Based on an ordinal scale each expert (P1 and P2) ranked the models’ predictions for 156 expert pre-defined ROIs for the categories “Segmentation”, “Shape Filter” and “Connective Tissue”, respectively. The absolute number and relative frequency of ROIs graded as “very bad” or “bad” are shown to the left of the zero line. The absolute number and relative frequency of predictions, which were ranked as “intermediate”, “good” or “very good” are visualized to the right of the zero line. Furthermore, to assess the significance of the ratings, for each category (“Segmentation”, “Shape Filter” and “Connective Tissue”) the p-values of a paired Student’s t-test are provided. In this context, the significance level was set to α = 0.05.
