AI-powered quantification of tumor-infiltrating lymphocytes from H&E stained images in ovarian cancer and its association with PARP inhibitor therapy outcomes.

Abstract

5575 Background: Maintenance therapy with PARP inhibitors improves the prognosis of patients with advanced or recurrent ovarian cancer. However, a simple biomarker detectable at treatment initiation remains unclear. This study evaluated the relationship between prognosis and artificial intelligence (AI)-powered quantification of immune cells in tumor and stroma areas in hematoxylin and eosin (H&E) slides. Methods: We evaluated 28 ovarian cancer patients treated with PARP inhibitors in our institution from 2019 to 2021. We developed an AI model to detect the epithelium and lymphocytes, specifically T cells (CD3+) and B cells (CD20+), from H&E-stained slides. This model was trained using annotated datasets, which included 26,509 images of the epithelium and 12,273 images of lymphocytes. The tumor bed areas were precisely defined by pathologists. Subsequently, the AI model identified epithelial and lymphocyte regions within these predefined areas. Using a treatment duration of PARP inhibitor with 12 months or more as the criterion, we identified the most relevant immune cell type based on the AUCs of the ROC curve and determined cutoff values. Furthermore, tumor BRCA1/2 status was assessed by a custom-targeted NGS testing panel. A log-rank test with a p-value < 0.05, considered statistically significant, examined the relationship between prognosis and tumor characteristics. Results: A total of 61 H&E slides were analyzed by AI: 28 slides before first-line chemotherapy, 25 after first-line chemotherapy, and 8 were the recurrent sample. The highest AUC was the lymphocyte-to-tumor area ratio (tumor-infiltrating lymphocyte score in tumor area, tTIL score) before the initial treatment, with an AUC of 0.73. The cutoff value of the tTIL score was set to 0.000534 based on the Youden index, dividing patients into tTIL-low (n=14) and tTIL-high (n=14) groups. The median follow-up period of the censored cases was 62.5 months vs. 77.3 months (p = 0.71). There were no significant differences in age, histological subtype, initial treatment method, proportion of R0 surgery, or that of tumor BRCA1/2 pathogenic mutations between groups. PARP inhibitors were used as maintenance therapy for recurrent settings in 71.4% of cases in both groups. The 5-year overall survival (5-y OS) rate of tTIL-high was significantly better than that of tTIL-low (84.4% vs. 30.8%, p = 0.0068). The median duration of OS after initiation of PARP inhibitors was significantly longer in the tTIL-high group (23.8 months vs. not reached, p = 0.046), and especially in tumor BRCA1/2-negative cases, tTIL-high had a significantly better 5-y OS rate of 90% vs. 12.5% (p = 0.0015). Conclusions: AI-powered tTIL score may predict the prognosis of ovarian cancer patients treated with PARP inhibitors. Future efforts will focus on increasing sample size and optimizing the tTIL score cutoff to improve accuracy.