A Pragmatic Machine Learning Approach to Quantify Tumor-Infiltrating Lymphocytes in Whole Slide Images.

Area of Science:

• Computational pathology
• Machine learning in oncology
• Immunohistochemistry analysis

Background:

• Elevated tumor-infiltrating lymphocytes (TILs) correlate with improved patient prognosis in various cancers.
• Manual quantification of TILs is subjective, time-consuming, and prone to inaccuracies.
• Accurate TIL assessment is crucial for predicting treatment response and patient outcomes.

Purpose of the Study:

• To develop and validate an automated computational method for quantifying TILs in hematoxylin and eosin (H&E)-stained lung cancer slides.
• To assess the correlation of computationally derived TIL levels with patient prognosis.
• To compare the performance of the automated method against existing state-of-the-art immune cell detection techniques.

Main Methods:

• Transfer learning of an open-source machine learning model (HoVer-Net) trained on public H&E slide data for nuclei segmentation and classification.
• Application of the model to quantify TILs in H&E whole slide images (WSIs) from lung cancer patients without manual data labeling.
• Prognostic correlation analysis using Cox proportional hazards models.

Main Results:

• The automated TIL quantification demonstrated a significant correlation with patient prognosis.
• HoVer-Net PanNuke Aug Model yielded a Hazard Ratio (HR) of 0.30 (95% CI 0.15-0.60) for patient survival.
• HoVer-Net MoNuSAC Aug model achieved an HR of 0.27 (95% CI 0.14-0.53), outperforming the current standard CD8+ T cell detection (HR 0.34).

Conclusions:

• Automated TIL quantification in H&E slides using a transferred machine learning model is a viable and accurate approach.
• This computational method offers a promising, efficient, and objective alternative to manual assessment for clinical use.
• Further validation is recommended prior to widespread clinical implementation.