Enabling DCIS subtyping: leveraging foundation models for robust grading and molecular biomarker scoring

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Breast cancer research

Background:

• Ductal Carcinoma In Situ (DCIS) is a precursor to invasive breast cancer, but lacks reliable prognostic markers, leading to overtreatment.
• Current management often involves intensive treatment for nearly all DCIS cases, even when unnecessary.
• The LORD trial aims to reduce overtreatment by offering active surveillance for specific DCIS subtypes.

Purpose of the Study:

• To develop and validate a deep learning pipeline for predicting ER, HER2, and grade status directly from H&E-stained digital pathology slides of DCIS.
• To support the LORD trial's objective of identifying eligible DCIS patients for active surveillance.
• To reduce overtreatment in DCIS management through accurate biomarker prediction.

Main Methods:

• A deep learning pipeline utilizing foundation models was developed to predict ER, HER2, and grade status.
• Models were trained and tested on a multicenter Dutch dataset (n=887) and externally validated on a UK dataset (n=259).
• Performance was evaluated using Area Under the Receiver Operating Characteristic Curve (AUROC), balanced accuracy, and Negative Predictive Value (NPV).

Main Results:

• The deep learning models achieved high AUROCs for ER (0.90 Dutch, 0.80 UK), HER2 (0.84 Dutch, 0.74 UK), and grade (0.86 Dutch, 0.75 UK).
• Stratification of patients for active surveillance criteria yielded balanced accuracies of 0.81 (Dutch) and 0.64 (UK).
• Negative Predictive Values (NPVs) for active surveillance eligibility were 0.86 (Dutch) and 0.76 (UK), indicating reliable patient selection.

Conclusions:

• The developed deep learning models demonstrate robust generalization across different cohorts for predicting key biomarkers in DCIS.
• These AI-powered tools reliably predict ER, HER2, and grade status from digitized H&E slides, aiding clinical decision-making.
• The findings support the use of these models to identify Ductal Carcinoma In Situ patients suitable for less aggressive management strategies like active surveillance.