Automated Molecular Subtyping of Breast Carcinoma Using Deep Learning Techniques

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Biomarker analysis

Background:

• Breast carcinoma molecular subtyping is crucial for prognosis and targeted therapy.
• Immunohistochemistry (IHC) analysis of ER, PR, HER2, and Ki67 is the standard method.
• Manual assessment is time-consuming and lacks automation.

Purpose of the Study:

• To develop a novel deep learning framework for automated molecular subtyping of breast cancer from IHC images.
• To enable accurate and efficient prediction of four key molecular biomarkers.

Main Methods:

• A modified LadderNet architecture was employed for segmenting immunopositive elements in IHC slides.
• The framework utilizes fully convolutional neural networks with adaptive learning pathways.
• Post-processing quantifies immunopositive elements for biomarker status prediction.

Main Results:

• The deep learning framework achieved high concordance with manual pathologist assessments for IHC biomarker evaluation.
• Segmentation models demonstrated robust performance for all four biomarkers (ER, PR, HER2, Ki67).
• Automated prediction of biomarker status was successfully validated.

Conclusions:

• The proposed automated method significantly speeds up breast cancer molecular subtyping.
• It reduces pathologist workload and associated costs.
• Facilitates timely targeted treatment for improved patient outcomes.