Automated machine learning-based classification corresponds to pathologist classification of head and neck cancer patient derived organoid morphology

Presentation: P004
Topic: Cancer Biology
Type: Poster
Authors: Anuraag S Parikh, MD1; Andres Klein-Szanto, MD, PhD2; Daria Vasilyeva, DDS1; Ogoegbunam Okolo, BS1; Victoria X Yu, MD1; Samuel P Flashner, PhD1; Cecilia Martin, BS1; Masataka Shimonosono, PhD1; Salvatore M Caruana, MD1; Scott H Troob, MD1; Angela J Yoon, DDS3; Devraj Basu, MD, PhD4; Fatemeh Momen-Heravi, DDS, PhD1; Hiroshi Nakagawa, MD, PhD1
Institution(s): 1Columbia University; 2Fox Chase Cancer Center; 3Medical University of South Carolina; 4University of Pennsylvania

Introduction: Head and neck squamous cell carcinoma (HNSCC) tumors demonstrate significant inter-tumoral variability in histologic features, including grade, degree of differentiation, patterns of invasion, and keratinization, as well as intra-tumoral variability in cellular atypia. Two-dimensional (2D) cell line models do not recapitulate these histologic features or demonstrate intercellular morphologic heterogeneity. Three-dimensional (3D) patient derived organoids (PDO) are alternative in vitro models that may better recapitulate primary tumor histology, but standardized methods to classify organoid histomorphology are lacking.

Methods: PDOs generated from 17 HNSCC tumors were assessed for histomorphologic features. H&E sections of organoids were scored from 1 (well-differentiated) to 4 (poorly differentiated) based on shape, layers of basaloid cells, nuclear atypia, and keratinization by a dedicated head and neck pathologist. Scanned H&E sections were also analyzed by the HALO image analysis software, following user training, for morphologic characteristics at the cellular and tissue level. Machine learning-based hierarchical clustering was then performed in R to place scored samples into four groups, and this clustering was compared with pathologist scoring, as well as pathologic characteristics of the original tumor tissue. Organoids and original tissues were also stained for Ki-67, phospho-gamma-H2AX, and p63 and characterized by proliferation and organoid formation rate.

Results: HNSCC organoids analyzed by a dedicated head and neck pathologist scored across the spectrum of well- to poorly differentiated in our 4-tiered scoring system. There was 82% concordance (p < 0,05, chi-square test) between pathologist-based scoring and automated classification by machine learning-based image analysis and clustering. Correlations between both pathologist and automated classification of organoid morphology and clinicopathologic features of original tumor specimens, including TNM stage, grade, perineural invasion, and lymphovascular invasion, were assessed. Immunohistochemical sections stained for Ki-67, phospho-gamma-H2AX, and p63 were scanned and analyzed by the image analysis software software, which enabled quantification of these markers at the single cell level. These three markers were variable across organoid samples and highly correlated in their staining patterns within samples. Organoids were also characterized by the heterogeneity in staining patterns across cells within each sample, as well as the degree to which staining correlated with staining in original tissues.

Conclusions: HNSCC organoids were reliably classified into four categories based on histomorphologic characteristics. Automated machine learning-based classification of organoid morphology showed excellent concordance with pathologist classification, suggesting machine learning approaches may reliably be used for such classification. A similar approach is also useful in quantifying molecular marker expression heterogeneity at the single cell level in HNSCC organoid samples.