Optimization of deep learning models for the prediction of gene mutations using unsupervised clustering

Area of Science:

• Digital pathology
• Computational biology
• Genomics

Background:

• Deep learning models are widely used for interpreting whole-slide images (WSIs) in digital pathology.
• Predicting genetic mutations from WSIs is a key application, with current models often assuming tumor regions hold the most predictive power.

Purpose of the Study:

• To develop an unsupervised clustering-based multiple-instance deep learning model for predicting genetic mutations from WSIs.
• To identify spatial regions within WSIs that are predictive of specific gene mutations.
• To investigate whether non-tumor tissues in the tumor microenvironment contribute to mutation prediction.

Main Methods:

• Utilized whole-slide images (WSIs) from three cancer types from The Cancer Genome Atlas.
• Developed an unsupervised clustering-based multiple-instance deep learning framework.
• The model identifies and excludes image patches lacking predictive information, focusing on relevant spatial regions.

Main Results:

• The proposed model achieved more accurate gene mutation predictions compared to models using all image patches and two published algorithms across three cancer types.
• Validated that relying solely on tumor regions for WSI-based mutation prediction may not yield optimal performance.
• Demonstrated that other tissue types within the tumor microenvironment can offer superior predictive ability than tumor tissues alone.

Conclusions:

• The study highlights the heterogeneity of the tumor microenvironment in predicting genetic mutations.
• Unsupervised clustering and identification of predictive image patches are crucial for enhancing deep learning models in digital pathology.
• Non-tumor tissues in the microenvironment can be as, or more, predictive than tumor regions for genetic mutations.