Fig. 1: Development of an IHC AI Model finetuned to detect PD-L1 CPS in gastric cancer biopsies.

a Gastric PD-L1 CPS AI Model. A multi-organ foundation IHC AI model was developed (Mindpeak, Hamburg, Germany) and adapted to detect PD-L1 expression by CPS methodology in GC/GEJC/EAC biopsy tissues. Pretraining of the neural network (NN) PD-L1 AI model employed 1.4 million annotations from 518 multi-organ biopsy cases sourced from 16 labs. Refinement of cell annotations was accomplished through construction of a PD-L1 CPS specific decision tree (Gastric Cell Atlas) to reconcile heterogenous histology and complex features. Gastric specific manual annotations were determined from 212 GC/GEJC/EAC biopsy cases (28-8 IHC stained WSIs) sourced from 11 labs, 7 scanners and inclusive of 406,867 GC/GEJC/EAC cells. Finetuning of the Gastric PD-L1 CPS AI Model was performed by using multi-head knowledge distillation. Patch analysis was employed stepwise to identify PD-L1 tissue image patches, then tumor segmentation (red invasive tumor, blue tumor-associated immune cells), then cell identification and finally cell classification. For CPS, single cell classification was defined for PD-L1 positive TC (posTC, red), PD-L1 negative TC (negTC, yellow) and PD-L1 positive IC (posIC, purple). Total tumor cells (TC) were calculated from posTC plus negTC counts. CPS equals the number of posTC, plus the number of posIC, divided by total viable TC, multiplied by 100 and in all cases was mathematically derived. WSI analysis by Gastric PD-L1 CPS AI Model as final output. b Gastric Cell Atlas. Three pathologists with deep expertise in PD-L1 staining interpretation reviewed 31 cases of GC/GEJC/EAC with challenging histology and complex cellular features including at least one of the following criteria: overlapping cells; faint, shared, and/or granular staining; presence of tumor associated mononuclear inflammatory cells; indiscernible tumor or non-tumor cells; presence or absence of tumor invasiveness; and accurate quantification of cell numbers. Decision criteria were devised for cell type, cell number and PD-L1 positivity. Principles from the Gastric Cell Atlas manual annotations were summarized and along with the Gastric Cell Atlas Decision Tree were provided to pathologists involved in the training phase of the Gastric PD-L1 AI model. c Gastric PD-L1 CPS AI Model Performance. Preliminary performance of the Gastric PD-L1 CPS AI Model was tested on 55 gastric cancer cases, independent from training set cases, sourced from 3 labs and 3 scanners and stained with PD-L1 IHC 28-8 pharmDX assay. The WSIs were reviewed by 2 expert pathologists and assigned a CPS ≥ 5 PD-L1 reference consensus score. Left, agreement of Gastric PD-L1 CPS AI Model against reference consensus score before Gastric Cell Atlas decision rules (81.8%) and after (96.4%). Right, Cohen’s Kappa before Gastric Cell Atlas decision rules (0.62) and after (0.91).