Breast Cancer Molecular Subtype Prediction on Pathological Images with Discriminative Patch Selection and...

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Computational pathology

Background:

• Accurate breast cancer molecular subtyping is crucial for personalized treatment.
• Tumor heterogeneity and sampling errors from single paraffin blocks can delay diagnosis and treatment.
• Current methods rely on immunohistochemistry (IHC), which can be costly and time-consuming.

Purpose of the Study:

• To develop an AI-driven method for predicting breast cancer molecular subtypes directly from Hematoxylin and Eosin (H&E) whole slide images (WSIs).
• To assist pathologists in selecting the optimal paraffin block for IHC, thereby reducing sampling errors and treatment delays.
• To address the challenges of WSI-level labels and noisy patches in weakly supervised learning.

Main Methods:

• A weakly supervised learning framework utilizing discriminative patch selection and multi-instance learning (MIL).
• Co-teaching strategy with two networks to learn representations and filter noise patches.
• Balanced sampling for dataset imbalance and a noise patch filtering algorithm using local outlier factor.
• A loss function integrating local patch and global slide information for fine-tuning the MIL framework.

Main Results:

• The proposed AI method effectively predicts molecular subtypes from H&E WSIs.
• The developed models demonstrated superior performance compared to senior pathologists.
• The approach successfully addresses challenges related to noisy patches and limited region information in WSIs.

Conclusions:

• AI-based molecular subtype prediction from H&E WSIs is a viable and effective approach.
• This method has the potential to significantly aid pathologists in pre-screening paraffin blocks for IHC.
• The AI tool can improve diagnostic efficiency and accuracy in breast cancer subtyping.