Fig. 5

Ablation experiments (number of annotated cases): whole-slide image (WSI) similarity analysis, comparison of segmentation accuracy for different approaches, and visual evaluation of segmentation accuracy. (A) Distribution of case-level similarity values for non-annotated WSIs (n = 200; against annotated slides) dependent on the number of used annotated WSIs. During this ablation experiment we evaluate the impact of the number of annotated WSIs included in training (TCGA colorectal cancer cohort). Conclusion: higher number of annotated images included results in higher similarity values for non-annotated slides. (B) Comparison of segmentation accuracy for different training approaches (TCGA dataset, annotated slide n = 5, 10 or 15; non-annotated slide n = 200). Three different approaches (supervised learning, traditional SSL, and SSL using S–o-M) are compared concerning pixel-wise segmentation accuracy (measured by Dice Score). Conclusions: new SSL S–o-M approach provides significantly better segmentation accuracies for both tumor and tumor stroma classes. In general, for all methods larger number of annotated slides results in higher accuracies; S–o-M approach achieves competitive accuracies with low number of annotated WSIs. (C) Visual comparison of multi-class segmentation accuracy in test slides for resulting final models in three approaches dependent on number of annotated slides used. The review of the images by pathology experts shows higher quality of segmentation for S–o-M approach, especially when lower number of annotated slides was used.