3464 - Machine Learning-Based Quantification and Classification of Fibroblasts in Gastrointestinal Cancer

Z. Zhang; Department of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai, China

Purpose/Objective(s): The prognostic value of cancer-associated fibroblasts is noticed. Owing to the functional heterogeneity, the classification of cancer-associatied fibroblasts lacks a standardized experimental method. For reducing the reliance on specialized imaging techniques and helping with prognostic stratification, we aimed to develop a machine learning model to automate and classify fibroblasts based on routine histopathology images. Materials/

Methods: A machine learning model was developed for detection, classification and quantitative evaluation of fibroblasts in 1424 hematoxylin-eosin-stained images from 573 gastrointestinal cancer patients. The computerized fibroblasts were investigated in association with overall survival. One public dataset was used for model training and internal validation. External validation was conducted on data from one institution.
Results: The machine learning model achieved high accuracies in detecting and classifying three different types of cells(tumor infiltrating lymphocytes: 88%, tumor cell: 79%, fibroblasts: 78%). Based on the extracted features, fibroblasts were categorized into two morphologically distinct classes. Class One fibroblasts are flat, elongated and lightly dyed, Class Two fibroblasts are small, thick and darkly dyed. The proportion of fibroblast subtypes stratified patients into two distinct risk groups across colorectal cancer and gastric cancer. Survival outcomes compared between groups: the proportion of Class One fibroblasts vs Class Two fibroblasts(for colorectal cancer: hazard ratio [HR], 0.61; 95% CI, 0.53-0.71; P < .050). Patients with a higher proportion of Class One fibroblasts had a significantly improved overall survival compared to those with a higher proportion of Class Two fibroblasts(for colorectal cancer: P < .001; for gastric cancer: P < .001).
Conclusion: In this study, we proposed a deep learning model for quantifying fibroblasts on H&E slides and classified fibroblasts morphologically into two classes. The proportion of fibroblast subtypes is a meaningful and independent prognostic marker and it could improve prognostic stratification.