Fig. 4

Ablation experiments (dataset): whole-slide image (WSI) similarity analysis, comparison of segmentation accuracy for different approaches, and visual evaluation of segmentation accuracy (A) Similarity analysis between annotated (n = 10) and non-annotated WSIs (n = 200). During this ablation experiment we use a different, single-center cohort of cases (UKK colorectal cancer cohort). The heatmap on the left side depicts the cross-slide case-level similarity values between annotated cases, where darker colors indicate higher similarity. The unrelated cases show higher similarity values compared to multi-centric TCGA cohort (Fig. 3A) due to similar cutting and staining quality. This difference is statistically significant based on a two-sample t-test (t = 10.38, p = 1.13 × 10⁻²²). The box plot on the right side illustrates the distribution of similarities for a larger cohort of non-annotated cases (n = 200). The non-annotated cases are grouped to the most similar case among annotated cases and the points are similarity values measured against these most similar cases. (B) Comparison of segmentation accuracy for different training approaches (UKK dataset, annotated slide n = 10, 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). We concentrate on tumor and tumor stroma classes as two bottleneck classes for multi-class tissue segmentation problem (explanation in Fig. 1 and Methods). The accompanying table (right side) provides the average Dice score with standard deviation of 3 independent experiments for each approach. Conclusion: For tumor class, new SSL S–o-M approach provides significantly better segmentation accuracies, statistical significance of the combined tumor and tumor stroma scores was assessed using a paired t-test (*** p < 0.001, ** p < 0.01); segmentation for tumor stroma class is better compared to supervised model and comparable to traditional SSL (with higher average Dice scores for S–o-M approach). (C) Visual comparison of segmentation accuracy/generated pseudo-labels in non-annotated slides for supervised model (as part of traditional SSL approach) and for most “similar” model (trained on the most similar annotated case). The review of the images by pathology experts shows higher quality of segmentation/pseudo-labels for S–o-M approach.