Contrastive Pre-Training and Multiple Instance Learning for Predicting Tumor Microsatellite Instability.

Area of Science:

• Computational pathology
• Machine learning in oncology
• Digital pathology

Background:

• Accurate classification of tumor MicroSatellite Stability (MSS) and Instability (MSI) is critical for gastrointestinal (GI) cancer prognosis and treatment decisions.
• Whole Slide Image (WSI) analysis offers a rich source of information for cancer diagnostics but requires advanced computational methods for accurate interpretation.

Purpose of the Study:

• To develop and validate a novel two-stage weakly supervised methodology for enhanced MSI prediction in GI cancers using WSI.
• To leverage the synergy of Multiple Instance Learning (MIL) and a Contrastive Clustering Network (CCNet) for improved MSI classification.

Main Methods:

• A two-stage weakly supervised framework integrating a contrastive learning-based feature extractor with MIL for efficient labeling.
• Development of a unique Contrastive Clustering Network (CCNet) for MSI prediction.
• Evaluation using colorectal cancer (CRC) and stomach adenocarcinoma (STAD) datasets, including experiments with transfer learning.

Main Results:

• The proposed methodology demonstrated notable improvements in MSI classification accuracy, outperforming existing methods.
• The framework showed efficacy and generalizability across CRC and STAD datasets.
• Transfer learning, particularly pretraining on STAD data and transferring to CRC data, yielded superior performance.

Conclusions:

• The developed weakly supervised MIL and CCNet framework represents an advance in computational pathology for GI cancer diagnostics.
• The findings highlight the potential for enhanced MSI prediction, aiding clinicians in personalized treatment strategies and improving patient outcomes.
• The study underscores the value of transfer learning in improving diagnostic accuracy for histopathological images.