Abstract PO-012: AI and digital pathology based on nucleus morphology for diagnosis, prognosis, and morphological-gene isolation Li Fraumeni as a model

• Published in: Clinical cancer research Vol. 27; no. 5_Supplement; p. PO-012
• Main Authors: Tsarfaty, Ilan, Iraqi, Ayman, Megides, Or, Cauzmer, Nizan, Shacham, Hila, Barokas, Doruk, Iraqi, Fuad, Kidron, Dvora
• Format: Journal Article
• Language: English
• Published: 01.03.2021
• ISSN: 1078-0432; 1557-3265
• DOI: 10.1158/1557-3265.ADI21-PO-012

Abstract Li–Fraumeni syndrome (LFS), also known as the sarcoma, breast, leukemia, and adrenal gland (SBLA) syndrome, is a rare, autosomal dominant, hereditary disorder that predisposes carriers to cancer development. An early diagnosis and prognosis of LFS and identification of a novel target gene are critically important so that affected families can seek appropriate genetic counseling as well as screening for early detection and personalized therapy of cancer. We created a p53 knockdown collaborative-cross (CC) model to isolate the nuclear morphology features' genes. CC consists of recombinant inbred lines generated by reciprocal crosses between numerous founder lines. The CC animal model is a reference to apply genetic traits mapping with high-resolution phenotypes such as digital pathology features. The tumors developed in the LFS-CC Model were extracted, embedded in paraffin, sectioned, stained with Hematoxylin-eosin, and subjected to whole-slide scanning. We developed a high-throughput infrastructure for nuclear segmentation as a tool to create new LFS tumor biomarkers. These novel biomarkers are based on extracted features from 1) classical image segmentation and feature calculation tools (QuPath). 2) convolutional neural networks (CNN) from a deep learning nuclei segmentation classifier. 3) combining the two approaches by calculating features with QuPath for the nuclei detected by the CNN. The k-mean grouping of the features extracted from the digital pathology shows mediastinal (90-95% accuracy) vs. non-mediastinal LFS-lymphoma (30% – 50%) diagnosis. K-mean grouping of the nuclear morphological features and survival analysis (Kaplan-Meier) demonstrate the prognosis capacity of the digital morphology groups (seven significantly different survival curcles with p-values ranging between <0.001 –0.0162). K-mean grouping of the features creates four morphological phenotypes for QTL (Quantitative Trait Locus) analysis. These analyses identify 12 loci in the chromosomes that code for new nuclear morphological candidate genes. We also developed a knowledge-based infrastructure to select candidate genes relevant to human lymphoma and sarcoma survival. These modifier genes can serve as targets for personalized therapy. These results demonstrate that digital pathology using a unique animal model can identify features that can serve as prognostic factors for LFS and help identify new targets for therapy that code for the alteration of the nuclear structure. These infrastructures can be utilized for other tumors in human research. Citation Format: Ilan Tsarfaty, Ayman Iraqi, Or Megides, Nizan Cauzmer, Hila Shacham, Doruk Barokas, Fuad Iraqi, Dvora Kidron. AI and digital pathology based on nucleus morphology for diagnosis, prognosis, and morphological-gene isolation Li Fraumeni as a model [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-012.