Abstract 593P
Background
Ovarian carcinoma is a leading driver of cancer-related mortality in women, but predicting long-term survival at diagnosis remains a challenge. We demonstrate that AI-powered pathology can classify cells and tissue regions in the high-grade serous carcinoma (HGSC) tumor microenvironment, and reveal nuclear morphology (NM) associated with patient outcomes, directly from digitized hematoxylin and eosin (H&E)-stained whole slide images (WSI).
Methods
Machine learning (ML) models were trained using H&E WSI of ovarian carcinoma samples from both publicly available (e.g. TCGA) and proprietary sources (Total N=352 slides from 257 subjects). ML models were trained to label tissue regions (cancer epithelium, stroma, necrosis) and cell types (cancer epithelial cells, fibroblasts, and inflammatory cells) using 138,632 pathologist-provided annotations. A separate nuclear segmentation model was deployed on H&E WSI from 93 subjects to characterize NM at single-pixel resolution. Survival and clinically relevant statuses were regressed vs. nuclear features using univariate Cox or logistic regressions. False discovery rate correction was applied.
Results
ML-based predictions of tissue and cell classes showed concordance with manual pathologist consensus labels. Analysis of NM revealed that greater inter-nuclear variation in staining intensity and minor axis length were associated with reduced overall survival, reflecting the effect of variability in chromatin characteristics (corrected p=0.023, Hazard Ratio (HR) = 2.51) and inter-nuclear heterogeneity in shape and size across a slide (corrected p=0.035, HR=2.11), respectively. Greater variation in nucleus minor axis length was also correlated with increased number of aneuploidy events involving gains in chromosome arms (Spearman correlation r=0.41, p<0.001), in turn predicting reduced overall survival (corrected p=0.04, HR=2.12).
Conclusions
This analysis reveals a relationship between cancer cell nuclear morphological diversity, aneuploidy, and patient prognosis. Digital pathology can therefore identify objective, quantitative biomarkers that may help pathologists better stratify patients at diagnosis, allowing clinicians to avoid potentially toxic therapy.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
PathAI, Inc.
Funding
PathAI, Inc.
Disclosure
C.M. Michener: Financial Interests, Personal, Stocks/Shares: MedaSync. C. Kirkup: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. B. Rahsepar: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. J.S. Iyer: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. J. Abel: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. K. Leidal: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. A. Khosla: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. B. Trotter: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. M. Lin: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. M. Resnick: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. B. Glass: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. I. Wapinski: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc. F. Najdawi: Financial Interests, Institutional, Full or part-time Employment: PathAI, Inc.