43/#512 Assessing robustness of an artificial intelligence derived histological biomarker across different sites of disease and in serial sections in tubo-ovarian high-grade serous carcinoma. (4th December 2022)

• Record Type: Journal Article
• Authors: Krishnan, Rayan; Tiu, Ekin; Krishna, Vrishab; Nimgaonkar, Vivek; Bhambhvani, Hriday; O'Donoghue, Odhran; Vrabac, Damir; Joshi, Anirudh; Liang, Brooke; Zhang, Xiaoming; Han, Lucy; Wang, Aihui; Krishna, Viswesh; Howitt, Brooke
• Is Part Of: International journal of gynecological cancer. Volume 32(2022)Supplement 3
• Page Start: A45
• Page End: A46
• Publication Date: 2022-12-04
• Subjects: Generative organs, Female -- Cancer -- Periodicals; 616.99465
• DOI: 10.1136/ijgc-2022-igcs.87
• Languages: English
• ISSNs: 1048-891X

Abstract

Abstract : Objectives: Histological biomarkers may produce different predictions for a single patient when using whole slide images of biopsies from different sites and even serial sections of the same tissue. Previous work had developed a signature of AI-derived morphologic features correlated with response to platinum-based chemotherapy in tubo-ovarian high-grade serous carcinoma (HGSC) specimens from The Cancer Genome Atlas (TCGA) (hazard ratio: 0.35). We aim to assess the robustness of this marker across different sites of disease and in serial sections. Methods: 489 sections from 10 tissue microarrays (TMA) corresponding to 44 patients with HGSC from Stanford Hospital were included in this study. Using the digitally scanned histologic images, we computed geometric features of nuclei extracted from tissue regions using segmentation models. TMA sections were stratified into low and high responder groups by the histologic signature previously associated with platinum-based chemotherapy response. Concordance (c-index (C)) and mean pairwise percent difference (MPPD) across all cores for a given patient were calculated to assess the robustness of the signature. Results: The prediction of the morphologic signature is consistent when computed across all cores/slides per patient (C:0.66, MPPD:30%). When stratified by site, the signature is similar across serial sections for samples from the ovary (C:0.71, MPPD:22%) and the omentum (C:0.70, MPPD:25%). The signature is also consistently robust irrespective of anatomic site (C:0.62, MPPD:26%). Conclusions: The artificial intelligence derived histological biomarker associated with response to platinum-based chemotherapy is generalizable across both ovarian and omental sites and consistent between serial sections in patients with HGSC.