TIL classification

TIL classification classifies tumor-infiltrating lymphocytes (TILs) in whole slide images (WSIs) using convolutional neural networks to map spatial TIL distribution and abundance across 23 cancer types.

Key Features:

• Deep Learning Workflow: Processes 50×50 µm tiled image patches represented as 100×100 pixels at 20× magnification for patch-level classification and mapping within gigapixel WSIs.
• Advanced CNN Architectures: Trains models using VGG16, Inception-V4, and ResNet-34 to distinguish TIL-positive and TIL-negative patches.
• Positive Classification Criterion: Defines TIL-positive patches based on the presence of two or more TILs.
• Comprehensive Training Dataset: Combines pathologist manual annotations (strong annotations) with computer-generated labels from a first-generation TIL model to reduce false positives and false negatives.
• Automated Thresholding: Converts model probabilistic outputs into binary TIL-positive/negative classifications to generate spatial TIL maps.
• Performance Improvements: Updated models report accuracy increases up to 13% and F-score increases up to 15% relative to prior iterations.
• TIL-Maps-23 Dataset: Includes a curated dataset named "TIL-Maps-23" comprising TIL maps for 7,983 WSIs spanning 23 cancer types.

Scientific Applications:

• Biomarker analysis: Enables analysis of TILs as biomarkers for predicting disease progression and clinical outcomes.
• Tumor microenvironment mapping: Supports comparative mapping of TIL distribution across diverse cancer types to study tumor immune microenvironments.
• Translational research and prognostication: Provides spatial TIL data useful for research into therapeutic strategies and patient prognostication.

Methodology:

Train VGG16, Inception-V4, and ResNet-34 CNN models on a dataset combining expert pathologist annotations and computer-generated labels; process 50×50 µm patches represented as 100×100 pixels at 20× magnification; classify patches as TIL-positive when containing two or more TILs; apply automated thresholding to convert probabilistic outputs into binary classifications and produce TIL maps.