The prognostic relevance of tumor-infiltrating lymphocytes (TILs) in non-small cell lung cancer (NSCLC) is well-established. However, manual TIL quantification in hematoxylin and eosin (H&E) whole slide images (WSIs) is laborious and prone to variability. To address this, we aim to develop and validate an automated computational pipeline for the quantification of TILs in WSIs of NSCLC. Such a solution in computational pathology can accelerate TIL evaluation, thereby standardizing the prognostication process and facilitating personalized treatment strategies.

We develop an end-to-end automated pipeline for TIL estimation in lung cancer WSIs by integrating a patch extraction approach based on hematoxylin component filtering with a machine learning-based patch classification and cell quantification method using the HoVer-Net model architecture. Additionally, we employ randomized patch sampling to further reduce the processed patch amount. We evaluate the effectiveness of the patch sampling procedure, the pipeline's ability to identify informative patches and computational efficiency, and the clinical value of produced scores using patient survival data.

Our pipeline demonstrates the ability to selectively process informative patches, achieving a balance between computational efficiency and prognostic integrity. The pipeline filtering excludes approximately 70% of all patch candidates. Further, only 5% of eligible patches are necessary to retain the pipeline's prognostic accuracy (c-index = 0.65), resulting in a linear reduction of the total computational time compared to the filtered patch subset analysis. The pipeline's TILs score has a strong association with patient survival and outperforms traditional CD8 immunohistochemical scoring (c-index = 0.59). Kaplan–Meier analysis further substantiates the TILs score's prognostic value.

This study introduces an automated pipeline for TIL evaluation in lung cancer WSIs, providing a prognostic tool with potential to improve personalized treatment in NSCLC. The pipeline's computational advances, particularly in reducing processing time, and clinical relevance demonstrate a step forward in computational pathology.