PD-L1 expression assessment in Angiosarcoma improves with artificial intelligence support

Figures

• 

  Fig. 1. Annotation process. The slide is analyzed (1), and tumor regions of interest (ROIs) are annotated WSI-wide (2). A small number of 512 × 512 patches were then taken from each investigated WSI to individually annotate the cell- and staining type (TC+, TC- or OC) in a subset of the patch's cells (3).

• 

  Fig. 2. CNN Training process. a) Shown is the number of patients and patches utilized to train the segmentation CNN to detect tumor regions, as well as the number of patients and annotated cells employed to train for the segmentation of cell types. b) Visualization of the deep learning pipeline: First, all 512 × 512 patches containing tumor regions are identified via a U-Net CNN (1). After this, cell instances in the identified patches are determined via Cellpose (2). Then, a U-Net predicts the cell type for each cell instance (3). The blue colored cells are TC-, while the yellow ones are OC. Finally, cell type counts are aggregated over all patches identified in (1). The tumor prediction score (4) is then calculated via division of the number of PD-L1 positive tumor cells by all tumor cells identified within the whole-slide image. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

• 

  Fig. 3. Tumor patch and cell type prediction accuracies. (A) Comparison of H&E and PD-L1 staining for precision, recall and Dice coefficient. The ROC curve further compares both methods indicating slightly better performance of PD-L1 staining. (B) Cell type classification performance through precision, recall and Dice coefficient visualized via bar chart and confusion matrix. The Dice coefficient is calculated through the actual semantic segmentation overlap. Error bars represent the standard deviations across three experimental runs.

• 

  Fig. 4. Scatterplot showing PD-L1 TPS assessment of the three pathologists, as well as the AI's prediction. Gray lines connect the individual PD-L1 expression predictions for each WSI.

• 

  Fig. 5. Distribution of over- and underscoring, as well as other causes leading the pathologists to update their PD-L1 assessment. These self-reported errors were identified in a secondary review process, where discrepancies between AI and pathologist prediction prompted a renewed assessment.

• 

  Fig. 6. PD-L1 Expression Evaluation Workflow. An AI-based method and a pathologist independently determine PD-L1 expression on a tumor slide, resulting in TC scores. Significant TC score variances trigger a detailed pathologist reassessment. The extended evaluation time enables score correction, illustrating how AI facilitates improved accuracy in time-sensitive diagnostic tasks.