Lee Cooper

Ph.D.

Northwestern University

Associate Professor

PLCO (Learn more about this study)

PLCOI-877

Dec 13, 2021

Breast cancer prognosis

Histology has been the gold standard for diagnosis of breast cancers, but until recently the role of histologic grading in prognostication and clinical stratification has been very limited. This is in part due to high interobserver variability in grading tubule formation, nuclear pleomorphism, and mitotic activity. The 8th edition of the AJCC Cancer Staging Manual now incorporates histologic grading, and histologic grade informs clinical decision-making. Advances in slide scanning microscopes and machine-learning (ML) now enable precise and quantitative characterization of tissue morphology, and using these methods to improve prognistication is a significant opportunity. ML can provide quantitative, objective, and reproducible assessments of characteristics like growth patterns, nuclear morphology, and immune infiltration that may contain latent prognostic information not captured by manual grading. We have developed an approach called the Histomic Prognostic Score (HPS) that is based on quantitative measurement of nuclear and tissue morphology, and that captures patterns in non-neoplastic elements such as stroma and tumor-infiltrating lymphocytes (TILs). We have developed these models on whole-slide images of HE stained formalin-fixed parafin-embedded (FFPE) tissues from The Cancer Genome Atlas and The Cancer Prevention Study (CPS)-II and CPS-3 cohorts from the American Cancer Society, and have shown that our HPS improves prognostic accuracy over manual grading. In this project we seek to further validate this model using the PLCO breast cancer cohort.

Aim 1. Apply our tools to whole-slide images of HE stained FFPE tissue sections from the PLCO breast cancer cohort to segment and classify tissue regions and cell nuclei, and to extract quantitative histology features from these images. We will visually assess the segmentation and classification quality and characterize systematic errors that are different from those observed in our development cohorts.

Aim 2. Compare histology feature distributions in the PLCO cohort to distributions in our development cohorts. We will use major criteria like histologic grade and stage to compare feature distributions between the PLCO and development cohorts. This more granular approach will control for differences in cohort composition and will allow us to identify biases in the quantitative features.

Aim 3. Validate our HPS model in the PLCO cohort. Using the features from Aim 1, we will test our HPS in the PLCO cohort. We will examine if reclassifications based on HPS improve prognostic accuracy over models based on manual histologic grading and other clinical features.