Finding Regions of Interest in Whole Slide Images Using Multiple Instance Learning

Area of Science:

• Digital Pathology
• Artificial Intelligence
• Computational Biology

Background:

• Whole Slide Images (WSI) are crucial in Digital Pathology but pose AI analysis challenges due to slide-level annotations.
• Pathology diagnostics and oncogene mutation data (e.g., TCGA) are often recorded at the specimen or slide level, not tile level.
• This creates a dual challenge: predicting cancer phenotype and linking cellular morphology to it.

Purpose of the Study:

• To address challenges in AI analysis of WSI with slide-level labels.
• To explore weakly supervised Multiple Instance Learning (MIL) for cancer phenotype prediction and mutation association.
• To investigate AI performance in tumor detection and TP53 mutation identification in Invasive Breast Carcinoma and Lung Squamous Cell Carcinoma.

Main Methods:

• Applied a weakly supervised Multiple Instance Learning (MIL) approach to Whole Slide Images (WSI).
• Utilized datasets from The Cancer Genome Atlas (TCGA) for Invasive Breast Carcinoma (TCGA-BRCA) and Lung Squamous Cell Carcinoma (TCGA-LUSC).
• Evaluated tumor detection and TP53 mutation prediction using novel additive MIL and Attention MIL architectures at various magnification levels.

Main Results:

• A novel additive MIL implementation achieved high performance (AUC 0.96), comparable to reference methods.
• Attention MIL slightly outperformed other methods (AUC 0.97).
• Different AI architectures showed distinct sensitivities to morphological features at various magnification levels, with TP53 mutation detection being most sensitive at higher magnifications.

Conclusions:

• Weakly supervised MIL is effective for analyzing Whole Slide Images (WSI) in Digital Pathology, even with slide-level annotations.
• AI models can predict cancer phenotypes and identify morphology-mutation associations.
• The study highlights the importance of magnification levels in AI-driven pathological analysis for molecular insights.