Outcome and Biomarker Supervised Deep Learning for Survival Prediction in Two Multicenter Breast Cancer Series

Area of Science:

• Computational pathology
• Machine learning in oncology
• Biomarker discovery

Background:

• Accurate prediction of clinical outcomes is crucial for guiding breast cancer diagnosis, treatment, and patient counseling.
• Existing prognostic models may benefit from integration of advanced computational methods and comprehensive biomarker data.
• Multicenter studies provide robust data for validating predictive models in diverse patient populations.

Purpose of the Study:

• To develop and evaluate a joint outcome and biomarker supervised multitask deep learning model for breast cancer patient outcome prediction.
• To integrate deep learning with expert knowledge for enhanced accuracy, robustness, and integrated prediction of breast cancer outcomes.
• To validate the model's generalization across different datasets and sample types (tissue microarray and whole-slide images).

Main Methods:

• Convolutional neural networks (CNNs) were trained on digitized hematoxylin-eosin-stained breast cancer tissue microarray (TMA) samples.
• The model used breast cancer-specific survival as the endpoint and was trained on the FinProg study cohort.
• The model was validated on independent test sets from the FinProg and FinHer multicenter studies, including whole-slide images, and combined with pathologist-assessed tissue characteristics.

Main Results:

• The multitask algorithm achieved a concordance index (c-index) of 0.59 on the FinProg test set and 0.57 on the FinHer validation set.
• The deep learning model proved to be a statistically independent predictor of survival, even after adjusting for established prognostic factors.
• Combining deep learning with pathologist-assessed tissue characteristics improved predictive accuracy to a c-index of 0.66.

Conclusions:

• A multitask deep learning algorithm, supervised by patient outcome and biomarker status, identified tissue morphology features predictive of survival in breast cancer.
• The developed algorithms demonstrated generalization capabilities across independent patient series and whole-slide breast cancer samples.
• This approach provides prognostic information that complements existing established prognostic factors, offering potential for improved clinical decision-making.