Immune subtyping of melanoma whole slide images using multiple instance learning

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology image analysis

Background:

• Melanoma patient outcomes depend on early prognostic markers and treatment stratification.
• Tumor transcriptome analysis identifies immune subgroups linked to survival but is costly and not routine.
• Current clinical workflows lack efficient methods for immune profiling of melanoma tumors.

Purpose of the Study:

• To develop and validate deep learning models for classifying melanoma immune subgroups using routine histopathology slides.
• To overcome the limitations of transcriptome analysis by utilizing readily available hematoxylin and eosin (H&E) stained slides.
• To establish a cost-effective and clinically applicable method for melanoma patient stratification.

Main Methods:

• Systematic assessment of six multiple instance learning (MIL) frameworks.
• Evaluation across five image resolutions and three feature extraction methods.
• Development of pathology-specific self-supervised models for gigapixel H&E slide classification.

Main Results:

• Pathology-specific self-supervised models at 10x resolution achieved superior immune subtype classification.
• Achieved mean AUC of 0.80 in a primary melanoma dataset and 0.82 in a TCGA dataset for 'high' vs 'low' immune classification.
• Successfully stratified patients into distinct survival groups (log rank test, P<0.005) based on immune status.

Conclusions:

• Deep learning on H&E slides offers a viable alternative to transcriptome analysis for melanoma immune profiling.
• MIL methods can identify novel biomarkers and assist clinicians in inferring tumor immune landscape.
• This approach enables patient stratification for improved melanoma treatment strategies without additional genetic testing.