PhiHER2: phenotype-informed weakly supervised model for HER2 status prediction from pathological images.

Area of Science:

• Computational pathology
• Biomedical imaging analysis
• Machine learning for cancer diagnostics

Background:

• Accurate Human Epidermal Growth Factor Receptor 2 (HER2) status identification is crucial for breast cancer (BC) prognosis and treatment.
• Pathological slides are the gold standard but analyzing high-resolution images with intratumoral heterogeneity is challenging.
• Computational analysis offers potential for discovering morphological patterns linked to HER2 status.

Purpose of the Study:

• To develop a phenotype-informed, weakly supervised multiple instance learning architecture (PhiHER2) for precise HER2 status prediction from BC pathological images.
• To leverage intratumoral morphological heterogeneity for improved prediction accuracy.
• To provide interpretable insights into morphological patterns associated with HER2 status.

Main Methods:

• Developed a hierarchical prototype clustering module to identify representative phenotypes in whole slide images (WSIs).
• Integrated phenotype embeddings into a cross-attention module for enhanced feature interaction and aggregation.
• Employed a phenotype-based feature space to capture and utilize morphological heterogeneity for HER2 prediction.

Main Results:

• PhiHER2 demonstrated superior WSI-level representation through phenotype guidance.
• The model significantly outperformed existing methods on real-world breast cancer datasets.
• Interpretability analyses provided explicit insights into the relationship between morphological phenotypes and HER2 status.

Conclusions:

• PhiHER2 offers a robust and accurate computational approach for HER2 status prediction in breast cancer.
• The phenotype-informed strategy effectively addresses the challenge of intratumoral heterogeneity in pathological images.
• The model's interpretability enhances clinical understanding of morphological drivers of HER2 status.