Artificial intelligence-powered spatial analysis of tumor microenvironment identifies immune phenotypes in H&E stained non-small cell lung cancer, colorectal cancer and urothelial cancer

Background Spatial distribution of inflammatory cells relative to tumor cells can identify patients likely to respond to immunotherapy or track responses over time. Development of a robust, reproducible method to determine tumor immune phenotypes (IP) would be a valuable tool for patient screening and efficacy mapping. A pathologist scoring method was previously developed and validated to identify desert, excluded and inflamed IPs in pan-CK/CD8 immunohistochemistry (IHC) stained non-small cell lung cancer (NSCLC, n=30), colorectal cancer (CRC, n=33) and urothelial cancer (UC, n=30) resection samples. We present a comparative analysis of H&E stained slides from this sample set using an artificial intelligence-powered Lunit SCOPE IO analysis to identify tumor IPs.

Methods Samples were stained with H&E and pan-CK/CD8 IHC. CD8 cells were quantified by image analysis in the tumor associated stroma (TAS) and tumor cell nests (TCN). Pathologists estimated the proportion of CD8 cell infiltration in TAS and TCN to obtain the IPs. In H&E WSIs from these samples, SCOPE IO quantified the spatial distribution of lymphocytes by histomorphology to determine case IPs.¹ Pathologist IPs in pan-CK/CD8 IHC stained WSI were compared to SCOPE IO IPs in H&E stained WSI from the same cases.

Results TCN and TAS area segmentation was highly concordant between SCOPE IO and pathologist pan-CK/CD8 IHC (spearman’s cc 0.91, 0.86). Both IP scoring methods are underpinned by regional cell density assessments relative to a cut-off. The excluded IP demonstrated best alignment with over two-thirds of SCOPE IO IP in agreement with pathologist assessments. A subset (25.0%) of pathologist desert cases were classified excluded by SCOPE IO, possibly due to algorithm detection of broader stromal lymphocytes populations. Inflamed versus excluded IP can be a source of discordance for pathologists and this was also apparent between pathologist and SCOPE IO with 32.4% of inflamed cases reported excluded by SCOPE IO, in part due to the low threshold for inflamed phenotype in the pathologist method. Adjustments to threshold criteria improved agreement between the pathologist and SCOPE IO methods for IP scoring.

Conclusions AI-powered Lunit SCOPE IO analysis of H&E stained slides is a promising tool for the identification of tumor immune phenotypes in carcinoma samples and would negate the requirement for IHC staining of epithelial or inflammatory cell markers. In the absence of a standard universal method for tumor IP assessment, insights gained from the comparative analysis of visual and AI-powered scoring serves to improve the robustness of both methods.