Abstract 39: Digital pathology foundation models enable accurate tumor organoid classification

Abstract This study develops an efficient and accurate pipeline for classifying tumor-derived organoid candidates as benign or malignant, addressing challenges posed by organoid culture and advancing personalized medicine applications. Background: Tumor-derived organoids replicate tumor architecture and molecular features, providing controlled environments to study tumor behavior and therapeutic responses. A key challenge in using organoid models is classifying cellular aggregates as benign or malignant. Benign cells in a tumor source can sometimes outcompete malignant tumor cells in culture, misrepresenting the tumor's characteristics. Also, histological characterization of organoids is challenging due to the absence of histologic clues from tissue organization. Current approaches, such as genomic sequencing, effectively resolve such uncertainties but are costly and time-intensive. Hypothesis: We hypothesized that a deep learning (DL) based computational image analysis classification pipeline could improve the accuracy and efficiency of distinguishing organoids as benign or malignant. Methods: We developed a novel pipeline to enhance organoid classification methods using ML. Our pipeline was developed and validated on H&E-stained slides obtained from organoids derived from 17 colorectal cancer patients (26 slides: 11 tumor, 15 normal) with various degrees of differentiation and paired normal controls. Ground truth data were established using a comprehensive genomic profiling panel to confirm that tumor organoids matched primary tumor features and, for tumors with high microsatellite instability, mismatch repair protein status. We first trained a segmentation model using a Human-in-the-Loop approach for precise identification of regions of interest. We extracted non overlapping 256x256 tiles at 20x magnification. We used the UNI foundation model to extract (k=1024) relevant features from each tile. We fitted Linear Discriminant Analysis (LDA) models on the UNI feature vectors using leave-one-patient-out cross validation according to patients. Each slide's final classification score was determined by averaging the LDA scores across all tiles from that slide. Results: The LDA classifier achieved 92.3% accuracy and AUROC of 0.964 (0.907-1.000), matching gastrointestinal pathologist performance. Additionally, we visualized the UNI feature vectors through UMAP plots. Plot evaluation demonstrates distinct clustering of unsupervised features into tumor and normal groups. Conclusions: These findings demonstrate the feasibility of coupling organoids with DL methods to facilitate organoid classification. Our approach offers an efficient new strategy for studying organoid morphology and advancing organoid models. Adaptations of this model are expected to be valuable in understanding organoid responses to various therapies, paving the way for improved personalized cancer care. Citation Format: Adi Orlyanchik, Matteo Sacco, James Dolezal, Joseph Kainov, Piao Zhao, Mohammed Aziz Khan, Marina Garassino, Marc Bissonnette, Le Shen, Alexander T. Pearson, Christopher Weber. Digital pathology foundation models enable accurate tumor organoid classification [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 39.