Artificial intelligence-powered tumor purity assessment from H&E whole slide images correlates with consensus purity estimation based on pathological examination and next-generation sequencing
Gahee Park, Sangjoon Choi, Seokhwi Kim, Soo Ick Cho, Wonkyung Jung, Jeongun Ryu, Minuk Ma, Donggeun Yoo, Kyunghyun Paeng, Chan-Young Ock, Sanghoon Song, Heon Song, Sergio Pereira, Seonwook Park
USCAP, 2022
Background: The advanced genomic analysis capturing tumor-specific profiles in DNA or RNA allowed us to estimate tumor purities more precisely. However, the direct evaluation of tumor purity assessment in a whole slide image (WSI) was limited due to technical challenges. Here, we present an artificial intelligence (AI)-powered tumor purity (AI-P) method using Lunit SCOPE IO to capture and quantify its tumor heterogeneity by defining characteristics of tissue area and cell types from WSI. Furthermore, we implemented AI-P for H&E stained WSI in The Cancer Genome Atlas (TCGA) cohorts to validate previous methodologies of next- generation sequencing (NGS)-based tumor purity estimates.
Design: Lunit SCOPE IO embedded two computer vision models: a tissue segmentation model and a cell detection model to analyze the spatial analysis of heterogeneity on WSI from the pan-cancer analysis. The performance of Lunit SCOPE IO was trained and validated with 3,166 H&E WSI of multiple cancer types, annotated by 52 board-certified pathologists. First, we implemented Lunit SCOPE IO to extract the characteristics of tissue and cell images from H&E stained WSI in TCGA cohorts. The AI-powered tumor purity (AI-P) was calculated as the total number of tumor cell counts in the cancer area over the total cell counts. Next, we evaluated AI-P and differential expression of tumor purity estimates from Aran et al., including breast cancer, lung cancer, and colorectal cancer of TCGA cohorts in 16 cancer types (n= 6,573). Finally, to assess clinical relevance driven by purity estimates in treatment effectiveness and prognosis, we classified AI P into “high” and “low” by a threshold of 50%.
Results: The overall average tissue size from the TCGA cohorts was 315.3 ± 165.6 (Mean ± SD) mm2 with the average cancer region of 65.9 ± 57.6 (Mean ± SD) mm2 which provided average AI-P estimates of 36.4% ± 18.7% (Mean ± SD). Ovarian cancer had the highest AI-P estimate of 51.2% amongst the 16 cancers. The purity estimates from AI-P and multiple genomic profiling methods (ABSOLUTE, ESTIMATE, and LUMP) had high concordance between most cancer types (|R| > 0.40, p < 0.001), but there was a loose correlation between AI-P and pathologic examination (|R| = 0.18, p < 0.001, Figure 1). Overall survival of the “high” AI- P group (n=1,663) prolonged significantly for those with above 50% of AI-P score compared to the “low” AI-P group (n=4,910, median OS: 85.1 versus 110.9 months, P=0.016). In a subgroup analysis, the ""high"" AI-P group had significantly poor prognosis in a total of 5 out of 16 cancers (p < 0.05) including kidney clear cell carcinoma (HR 2.26, p = 0.006), uterine corpus endometrial carcinoma (HR 1.71, p = 0.032), colorectal cancer (HR 1.67, p = 0.009), breast cancer (HR 1.47, p = 0.021), and hepatocellular carcinoma (HR 1.45, p = 0.037, Figure 2).
Conclusions: We demonstrated that AI-P provides compatibility of quantifying tumor purity throughout the comparative analysis of genomic-profiled tumor purity measurements. Thus, we believe that using AI-P from Lunit SCOPE IO can practically and expeditiously assess the tumor purity from H&E slides.