Deep learning based quantitative models to distinguish tumoral and stromal tumor-infiltrating lymphocytes and the prognosis patients with high-grade serous ovarian cancer.

Abstract

84 Background: The better prognostic value of tumor-infiltrating lymphocytes (TILs) has been reported in high-grade serous ovarian cancer (HGSOC). However, the TILs evaluation of previous reports largely depends on manual quantification based on a visual assessment within limited section of pathological slides or tissue microarrays, which is subject to inter-observer variability. Since tumor immune microenvironment (TIME) of HGSOC is highly heterogeneous, there is a need to quantify TILs in large areas of pathological slides in an objective and time-saving manner. Here we present an AI-powered quantification method to distinguish tumoral TILs and stromal TILs using hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) of HGSOC. Methods: 161 H&E stained WSIs of HGSOC patients treated at our hospital between 1998 and 2021 were prepared, and by using these WSIs we generated two AI-powered models. Initially, a TILs detection model (TIL-model) was created using a publicly available dataset of 86104 pan-cancer TILs images of 100 × 100 pixels at magnification of x200 (Saltz et al., 2018, Cell Reports). Secondary, a tumor and stroma segmentation model (T/S seg-model) was developed using 136962 patches of 224 × 224 pixels at magnification of x50 which were manually annotated with tumoral regions and stromal regions in 102 H&E stained WSIs of HGSOC from The Cancer Genome Atlas (TCGA). Deep convolutional neural networks of NASNet-Large for the TIL-model and TransUNet for the T/S seg-model were used. TILs density score (number of TILs positive tiles per tumor area) of each sample was calculated, and association between TILs density score and the prognosis was investigated. Results: The established models achieved high accuracy: the TIL-model showed an area under the curve (AUC) value of 0.896, and the T/S seg-model showed an intersection over union value of 0.967. By combining the two models, we successfully distinguished tumoral TILs and stromal TILs. Using our dataset of 161 WSIs, we split the patients by TILs density score at top quartile into high tumoral TILs subgroup (n=40) and low tumoral TILs subgroup (n=121). High tumoral TILs subgroup showed significantly longer overall survival (OS) and progression free survival (PFS) (OS: median 70 months vs not reached, p < 0.005. PFS: median 34 months vs 21 months, p = 0.02). Conclusions: We developed deep learning based pathological AI models to detect TILs and tumors using H&E stained WSIs, which achieved to quantify tumoral and stromal TILs. The objective assessment of TILs using these models will lead to the development of clinically applicable tools for the evaluation of the immune status in HGSOC.