Abstract 94P
Background
Programmed death-ligand 1 (PD-L1) expression is a predictive marker for immune checkpoint inhibitors (ICI) treatment in urothelial carcinoma (UC). The combined positive score (CPS) is a method to evaluate the expression level of PD-L1 in UC. This study aimed to evaluate the performance of an artificial intelligence (AI)-powered PD-L1 CPS analyzer on UC compared to the pathologists.
Methods
Lunit SCOPE PD-L1 CPS was developed with 3.02 x 105 tumor cells and 3.49 x 105 immune cells from PD-L1 immunohistochemistry-stained whole-slide images (WSI) of UC from multiple institutions, annotated by 94 pathologists. The tissue area segmentation and cell detection AI models were developed based on a semantic segmentation algorithm, which includes an atrous spatial pyramid pooling block. To validate the model, a total of 543 PD-L1 stained UC WSIs were obtained from three university hospitals (n = 245, 205, and 93 in each). Three uropathologists evaluated slide-level CPS and assigned CPS high or low (10% cutoff value). The agreement (high or low) or correlation (continuous value) of CPS between the pathologists and the AI prediction was evaluated.
Results
All pathologists agreed on the CPS level in 446 out of 543 cases (82.1%). The agreement or correlation between either of the two pathologists was 87.1%−89.9% (Table). AI model accuracy to pathologists' consensus was 88.8%, and the Intraclass correlation coefficient (ICC) value between the AI model CPS value and the average CPS value of pathologists was 0.94 (95% confidence interval [CI] 0.93−0.95). The performance of the AI model was similar with each individual pathologist (accuracy / ICC; 85.1% / 0.94, 86.6% / 0.90, and 87.1% / 0.93, respectively) and individual hospital dataset (accuracy / ICC; 89.4% / 0.92, 87.8% / 0.95, and 89.2% / 0.92, respectively). Table: 94P
| Comparison | Agreement (10% cutoff) | Correlation (ICC (2,k)) (95% CI) |
| Pathologist A vs. B | 87.1% | 0.91 (0.90−0.93) |
| Pathologist A vs. C | 89.9% | 0.96 (0.96−0.97) |
| Pathologist B vs. C | 87.3% | 0.92 (0.90−0.93) |
| Pathologist consensus vs. AI model | 88.8% | 0.94 (0.93−0.95) |
Conclusions
This study demonstrates that an AI-powered PD-L1 CPS analyzer can classify CPS levels in UC comparable to pathologists.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Lunit.
Funding
Lunit.
Disclosure
S.I. Cho: Financial Interests, Personal, Stocks/Shares: Lunit; Financial Interests, Personal, Full or part-time Employment: Lunit. W. Jung: Financial Interests, Personal, Stocks/Shares: Lunit; Financial Interests, Personal, Full or part-time Employment: Lunit. S. Kim: Financial Interests, Personal, Full or part-time Employment: Lunit. G. Park: Financial Interests, Personal, Stocks/Shares: Lunit; Financial Interests, Personal, Full or part-time Employment: Lunit. S. Song: Financial Interests, Personal, Stocks/Shares: Lunit; Financial Interests, Personal, Full or part-time Employment: Lunit. C. Kang: Financial Interests, Personal, Full or part-time Employment: Lunit. M. Ma: Financial Interests, Personal, Full or part-time Employment: Lunit; Financial Interests, Personal, Stocks/Shares: Lunit. D. Yoo: Financial Interests, Personal, Stocks/Shares: Lunit; Financial Interests, Personal, Full or part-time Employment: Lunit. K. Paeng: Financial Interests, Personal, Stocks/Shares: Lunit; Financial Interests, Personal, Full or part-time Employment: Lunit. C. Ock: Financial Interests, Personal, Stocks/Shares: Lunit, Ybiologics; Financial Interests, Personal, Full or part-time Employment: Lunit; Financial Interests, Personal, Member of the Board of Directors: Ybiologics. All other authors have declared no conflicts of interest.