Machine Learning to Refine Breast Staging for Prediction of Axillary Node Status and Prognosis

Principal Investigator

Name
Mark Sherman

Degrees
M.D.

Institution
Mayo Clinic, Jacksonville, Florida

Position Title
Professor of Epidemiology and Laboratory Medicine and Pathology

Email
sherman.mark@mayo.edu

About this CDAS Project

Study
PLCO (Learn more about this study)

Project ID
PLCOI-2066

Initial CDAS Request Approval
Aug 10, 2026

Title
Machine Learning to Refine Breast Staging for Prediction of Axillary Node Status and Prognosis

Summary
Breast cancer (BC) size, represented by “T” in TNM staging, is important for defining prognosis and management. T is defined by gross measurement; however, BC masses of equal sizes contain highly variable numbers of invasive BC cells and differ in relative amounts of carcinoma in-situ, benign breast disease, stroma, inflammation, other cell types and structures. Thus, we hypothesize that adjustment of T stage by measurements that reflect the burden of invasive BC cells could improve prediction of prognosis; however, this has not been tested. We further propose that evaluating the distribution of invasive BC cells within a mass may reflect the degree of “dispersion” or “invasiveness” and thus influence clinical behavior. Leveraging use of digital images of H&E-stained sections for pathology diagnosis, we aim to refine an automated machine learning (ML) algorithm that reflects the invasive BC burden and can be used to adjust the T-value to improve staging and guide precision therapy and prognosis. In future work, we are extending this effort by analyzing multiplex phenotyping of individual BC cells, yielding a novel transformational approach for prognosis and guiding precision therapy. For example, we aim to provide a measure that reflects that burden of ER+/Ki67+ BC cells in a BC by adjusting the BC size, measured grossly (T stage), by invasive BC metrics and further by percentage of labeled cell types. To that end, we trained 2 segmentation models on 776 annotated tiles in 110 whole slide H&E images (WSIs) at resolutions of 0.5 um per pixel and 2 um per pixel. Pixel-wise predicted probabilities from these models are averaged to yield a probability score for invasive BC per pixel. Currently, the dice scores for these models are 0.82 and 0.85 (Dice Score reflects strong agreement with pathologist's labels). Next, we developed and preliminarily evaluated a five-step pipeline to assess BC burden: invasive cancer segmentation, CellViT-based cell detection, cancer segmentation correction using the cell detection results, tumor mass estimation, and circumscription calculation. The BC mass area on a slide is based on stromal adipose tissue boundaries, which we propose as representative for estimating cell composition; however, we use the BC size determined clinically in our model. We consider invasive BC area and invasive BC number within the mass as numerators, the estimated BC mass area as the denominator and compute ratios. To evaluate the algorithm, we are collaborating with the American Cancer Society (ACS) cohorts which provided us with nearly 4,000 images of BCs with detailed risk factors and follow-up. Our sample includes invasive BC, not treated with neoadjuvant therapy or ablation and without metastases at initial presentation. Any surgical procedure, including bilateral mastectomy, was deemed acceptable. Images with interfering artifacts and small biopsies were excluded. We split our sample into discovery and test sets. We applied the algorithm to a discovery set of images and found that adjusting BC size (T stage) for these features improves prediction of axillary node status and survival. Evaluation of the test set is being completed.

Aims

• Aim 1: To develop and refine a machine learning algorithm to assess features in H&E scanned images that predict axillary lymph node status and survival. Proposed features under refinement include: 1) delineation of BC masses within H&E images by defining BC vs. non-BC boundaries; 2) mapping BC areas within delineated BC masses; 3) estimation of invasive BC cells within delineated masses and 4) Quantification of invasive cancer spatial distribution within the tumor mass (e.g., localized versus diffusely dispersed tumor growth). We have shown that these metrics can refine prediction of axillary node status and survival based on BC size (i.e., T stage) in a discovery set using the ACS cohort data. Test set is under evaluation. We are assessing various covariates and strata, including BC risk factors and clinical factors, such age, menopausal status, body mass index, smoking, alcohol use, grade, estrogen receptor (ER) status (only marker available in ACS cohort) and others.
• Aim 2: To confirm performance of machine learning algorithm in predicting axillary lymph node status and survival in an independent cohort, the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial. The ACS cohort offers a strong platform for machine learning development because women are recruited throughout the U.S., providing great patient diversity and variation in specimen handling and preparation. Thus, analyses in this resource may limit risks of overfitting that might be problematic in single institution or pathology laboratory studies. However, images are only available for a subset of BC patients in the cohort and data including follow-up is provided mainly by participants. Thus, confirmation of our findings in another diverse cohort is important. Furthermore, PLCO offers a rich data source that may complement those available in ACS, especially HER2 testing of BCs, biological serum measurements and long-term follow-up with mortality data, which is particularly important for ER-positive BCs that may recur after many years. While we knowledge that some treatments in PLCO may vary from current practice, reduced reliance on neoadjuvant therapy during the time period of the study and the enrollment at academic centers suggests that many scanned images are derived from well-handled surgical pathology specimens ideally suited to this project.

Collaborators

Mark Sherman Mayo Clinic, Jacksonville, Florida
Peiliang Lou, PhD Mayo Clinic, Jacksonville, Florida
Stacey Winham, PhD Mayo Clinic, Jacksonville, Florida
Laura Pacheco-Spann, DHSc Mayo Clinic, Jacksonville, Florida
Lauren Teras, PhD American Cancer Society
Lee Cooper, PhD Northwestern University