Clara Bodelon

Ph.D.

American Cancer Society

Senior Principal Scientist

PLCO (Learn more about this study)

PLCOI-1947

Jul 11, 2025

Digital Pathology Prognostic of Late Breast Cancer Outcomes

About 40,000 breast cancer deaths occur annually in the US, two-thirds of which are among
patients originally diagnosed with non-metastatic disease. These deaths are preceded by a
recurrence. While almost all recurrences among women with ER-negative disease occur in the
first 5 years, women with ER-positive disease (~74% of breast cancers) have a substantial
recurrence hazard long after 5 years. In 2014 there were 3.1 million breast cancer survivors in
the US—two-thirds of whom had survived at least 5 years—and nearly 4 million breast cancer
survivors are expected by 2024. With improved breast cancer prognosis, the proportion of
recurrences and deaths occurring 5 or more years after diagnosis has steadily increased. It is
now estimated that about 1 in 3 breast cancer recurrences occur more than 5 years after
diagnosis. Existing and emerging therapies are available to prophylactically target late
recurrences, but associated toxicities require that they be used only for high-risk women.
However, there is no accepted late recurrence risk stratification model to guide their use. NCCN and ASCO have both said in their guidelines for biomarker-guided therapies that no established prognostic marker predicts late recurrence risk in ER-positive breast cancer patients. ASCO’s latest guidance explicitly states: “Well-conducted studies specifically designed to establish prognostic markers of late recurrence are needed.”

The proposed project will fill this important evidence gap. We aim to use artificial
intelligence (AI) methods to identify latent characteristics of digital pathology images of the first
breast cancer that specifically predict high risk of late breast cancer death (as a proxy for late recurrence). This prognostic marker will have immediate clinical utility. Almost all non-metastatic breast cancers are initially treated by a surgery that excises the tumor and the excised tissue is dissected and stained. In recent years, many of the stained slides are digitally scanned and stored. Formalin-fixed paraffin embedded slides are routinely archived for long duration after surgery, so can be recovered and digitally scanned. Thus an AI-generated algorithm that is prognostic of late recurrence could be applied to a high proportion of the ~4 million US 5+ year survivors of breast cancer.

Using American Cancer Society (ACS) Cancer Prevention Study-II (CPS-II) participants who developed breast cancer during cohort follow-up, and for whom digital pathology images are available, we will use image analysis and AI techniques to comprehensively characterize breast cancer tumor cell morphology, immune cell architecture, cell interactions, and intra-tumor heterogeneity from digitized H&E-stained specimens. Using metrics derived from these digital images, we will train AI models to identify patients with breast cancer at higher risk of late breast cancer-specific mortality by dividing the cohort into three survival periods: [0–5 years, 5–10 years, 10–25 years].

Aim: Validate the model predictive of late breast cancer mortality developed in the ACS CPS-II study in the PLCO cohort using existing digital imaging and breast cancer survival data.