2 Clarke Drive
Cranbury, NJ 08512
© 2022 MJH Life Sciences™ and OncLive - Clinical Oncology News, Cancer Expert Insights. All rights reserved.
December 9, 2020 - The artificial intelligence–digital breast cancer risk discrimination platform PreciseDx was able to classify patients with Oncotype Dx low-risk recurrence scores with high accuracy using only hematoxylin and eosin stain images and limited clinical data.
The artificial intelligence (AI)–digital breast cancer risk discrimination platform PreciseDx was able to classify patients with Oncotype Dx low-risk recurrence scores with high accuracy using only hematoxylin and eosin stain (H&E) images and limited clinical data, according to findings from a study presented during the 2020 San Antonio Breast Cancer Symposium.1
Results showed that the PreciseDx model, which combined factors such as age and progesterone receptor (PR) levels with imaging features indicative of mitotic activity and nuclear characteristics, was able to determine low-risk recurrence scores with a positive predictive value of 91% (95% CI, 0.87-0.94) and a sensitivity of 92% (95% CI, 0.86-0.95).
Additionally, the model was able to these scores with a positive/negative likelihood ratio of 6.8 and 0.5, respectively. PreciseDx utilized 12 imaging features that represented nuclear morphology, gland morphology, and mitotic activity, such as automated and quantitative grade.
“Our goal is to provide a precise mathematical approach to classifying and treating disease, which assists our clinicians with information for effective patient care and health management,” Carlos Cordon-Cardo, MD, PhD, director of the Center for Computational and Systems Pathology and Precise Diagnostics, chair of the Department of Pathology, and professor of Pathology, Genetics and Genomic Sciences, and Oncological Sciences at Mount Sinai’s Icahn School of Medicine, commented in a past press release on the platform.2“By refining diagnoses, we can save patients from unnecessary treatments and improve outcomes.”
Clinical breast cancer guidelines have underscored the significance of grading and staging in breast cancer to inform treatment decisions. Although histologic grade is a subjective factor that is known to be non-quantitative, skill-dependent, and sometimes inaccurate, it continues to serve as an independent prognostic feature; as such, it is a key player in the management of patients with this disease. Investigators designed PreciseDx to predict breast cancer recurrence through the combined utilization of digital H&E features relevant to grading and select clinical variables.
The study, which was shared as a poster presentation during the meeting, looked at a cohort of 391 patients with breast cancer and invasive ductal carcinoma who had at least 1 conventional H&E image; the data for this cohort of patients were collected from Mount Sinai Hospital. The data collected from these patients comprised clinical features, such as age, hormone receptor status, stage, and tumor size, as well as Oncotype Dx recurrence scores.
Patients included in the study had a mean age of 57 years, and all had stage I/II disease. More than half (59%) of the patients had grade 2 disease and 6% were positive for involvement of 0-3 lymph nodes. Additionally, 97% of the patients had invasive ductal carcinoma, all had estrogen receptor–positive disease, and 94% were PR positive. No patients were reported to have HER2 amplifications.
The study had a median follow-up of 61 months. Notably, 323 patients (83%) were considered to have low risk recurrence scores (</=25) while 68 were considered to have high risk recurrence scores (>25). Events occurred in 6% of patients (n = 23), half of which were locoregional recurrence (56%; n = 13).
Investigators utilized multiple deep learning strategies during the study to identify tumor regions and characterize cellular attributes, including overall gland structure, nuclear morphology, and mitotic figures. From there, features were produced and used to classify both high and low risk in patients, as defined by Oncotype Dx with automatic feature selection.
According to investigators, future models will extend outcomes to 10 years and examine both treatment selection and duration.