AbstractWithin the tumor microenvironment, specifically aligned collagen has been shown to stimulate tumor progression by directing the migration of metastatic cells along its structural framework. Tumor-associated collagen signatures (TACS) have been linked to breast cancer patient outcome. Robust and affordable methods for assessing biological information contained in collagen architecture need to be developed. We have developed a novel artificial neural network (ANN) based approach for tumor collagen segmentation from bright-field histology images and have tested it on a set of tissue microarray sections from early hormone receptor-positive invasive ductal breast carcinoma stained with Sirius Red (1 core per patient, n = 92). We designed and trained ANNs on sets of differently annotated image patches to segment collagen fibers and extracted 37 features of collagen fiber morphometry, density, orientation, texture, and fractal characteristics in the entire cohort. Independent instances of ANN models trained on highly differing annotations produced reasonably concordant collagen segmentation masks and allowed reliable prognostic Cox regression models (with likelihood ratios 14.11–22.99, at p-value < 0.05) superior to conventional clinical parameters (size of the primary tumor (T), regional lymph node status (N), histological grade (G), and patient age). Additionally, we noted statistically significant differences of collagen features between tumor grade groups, and the factor analysis revealed features resembling the TACS concept. Our proposed method offers collagen framework segmentation from bright-field histology images and provides novel image-based features for better breast cancer patient prognostication.
Tumor collagen framework from bright-field histology images predicts overall survival of breast carcinoma patients
