Multi-Scale Multiple Instance Learning for Lymph Node Metastasis Prediction in Early Gastric Cancer.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Lymph node metastasis (LNM) is crucial in gastric cancer progression.
• Multiple instance learning (MIL) is used for whole slide image (WSI) analysis.
• Current MIL methods often lack multi-scale analysis, unlike human pathologists.

Purpose of the Study:

• To develop a cross-scale multiple instance learning (MIL) framework for predicting lymph node metastasis (LNM) in early-stage gastric cancer (EGC) whole slide images (WSIs).
• To incorporate multi-scale interactions for improved diagnostic accuracy.
• To enhance the interpretability of automated LNM diagnosis.

Main Methods:

• Proposed a novel cross-scale MIL framework with multi-scale interactions.
• Introduced a cross-scale attention module to integrate features from different image resolutions.
• Aggregated cross-scale and resolution-specific features for slide-level LNM prediction.
• Validated the model on a clinical cohort of 740 T1-stage gastric cancer WSIs.

Main Results:

• The proposed framework achieved an Area Under the Curve (AUC) of 0.712.
• Outperformed baseline MIL models in predicting LNM.
• Generated multi-scale attention visualizations for enhanced interpretability.

Conclusions:

• The developed deep learning model accurately predicts LNM from EGC WSIs, offering an alternative to analyzing lymph node specimens.
• This approach can aid in treatment decisions for EGC patients, potentially preventing unnecessary lymph node resection.
• The study highlights the benefit of multi-scale analysis in computational pathology for cancer diagnosis.