ARTIFICIAL intelligence models accurately predicted key biomarkers in ductal carcinoma in situ directly from digitised pathology slides, according to new research aiming to improve selection for active surveillance and reduce overtreatment.
Ductal carcinoma in situ (DCIS) is recognised as a non-obligate precursor of invasive breast cancer. However, because reliable prognostic markers remain limited, most women diagnosed with DCIS undergo intensive treatment regardless of individual progression risk. Investigators developed a deep learning pipeline to identify patients who may be eligible for less aggressive management within the context of the LORD trial.
The study focused on women with screen detected, oestrogen receptor positive, human epidermal growth factor receptor 2 negative, grade 1 or 2 DCIS. Researchers trained and tested the artificial intelligence models using haematoxylin and eosin-stained digitised pathology slides from a Dutch multicentre dataset including 887 patients. External validation was then performed using an independent United Kingdom dataset of 259 patients.
Deep Learning Models Predicted Key DCIS Biomarkers
The models were designed to predict tumour grade, oestrogen receptor status, and human epidermal growth factor receptor 2 status directly from pathology images. In the Dutch dataset, model performance was high across all biomarkers, with mean area under the receiver operating characteristic curve values of 0.90 for oestrogen receptor status, 0.84 for human epidermal growth factor receptor 2 status, and 0.86 for grade prediction.
External validation demonstrated lower but still clinically relevant performance, with area under the receiver operating characteristic curve values of 0.80 for oestrogen receptor status, 0.74 for human epidermal growth factor receptor 2 status, and 0.75 for grade.
DCIS Surveillance Eligibility Identified Using AI
Researchers combined model outputs to stratify patients according to active surveillance eligibility criteria. This approach achieved balanced accuracies of 0.81 in the Dutch cohort and 0.64 in the United Kingdom cohort.
Negative predictive values were 0.86 and 0.76 for the Dutch and United Kingdom datasets respectively, suggesting the models may help identify patients unlikely to require intensive intervention.
Potential to Reduce Overtreatment
The findings support the potential use of artificial intelligence assisted pathology tools to guide treatment decisions in DCIS. Investigators concluded that the models generalised across cohorts and reliably predicted clinically relevant biomarkers associated with eligibility for active surveillance.
The authors suggested these approaches may ultimately support less aggressive management strategies for selected patients with DCIS, potentially reducing unnecessary treatment while maintaining patient safety.
Reference
Doyle S et al. Enabling DCIS subtyping: leveraging foundation models for robust grading and molecular biomarker scoring. Npj Breast Cancer. 2026; https://doi.org/10.1038/s41523-026-00957-6.
Featured image: Siam on Adobe Stock.
