TIL Pattern Identification and Correlation from Application of Deep Learning to Pathology Images

Pathologists traditionally make a diagnosis by examining patient specimens on glass slides under a microscope. Advances in whole slide imaging (WSI) allow pathologists to view the entire microscope slide as a digital image. These high-resolution digital images can be computationally analyzed using a number of techniques including the application of artificial intelligence (AI). Deep learning is a branch of AI that computationally mimics the way layers of neurons learn in the human brain. Deep learning in pathology imaging allows software to learn to recognize patterns in digital images that may represent disease or correlate with patient outcome.

The National Institutes of Health (NIH) have recently completed the Pan-Cancer Atlas, an in-depth analysis of molecular and clinical information from over 10,000 tumors representing 33 types of cancer. This valuable data set has been published as a collection of over 27 papers. One of these manuscripts was recently published in Cell Reports describing a ‘Computational Stain’ of tumor-infiltrating lymphocyte (TIL) patterns from digitized H&E-stained images of over 5,000 samples from 13 different cancer types. Importantly, the results from this approach were linked to patient clinical outcome and highlighted the imminent role of deep learning applied to whole slide imaging techniques and its direct impact on patient care.

Here we present a brief overview of some main points, but we highly suggest you read the open access article in its entirety.

Key Take Aways

  • In April 2017, the FDA approved the use of whole-slide imaging for primary diagnostic use.
  • For the first time a digital, deep-learning approach of whole slide imaging was used to quantify spatial characterization of TILs of 13 different cancer types.
  • TIL densities and spatial structure were found to be differentially enriched among tumor types, immune subtypes, and tumor molecular subtypes.
  • These TIL patterns were linked to genomic characterization and, collectively, the results were correlated with patient clinical outcomes.
  • This study indicates that quantification of diagnostic tissue may be used to inform the direction of research and/or diagnostic and prognostic indicators for immuno-oncology therapy.

Computational Pathology Method

  1. 5,202 digital whole slide images from 13 different cancer types were obtained from The Cancer Genome Atlas (TCGA).
  2. Systematic employment of automated image processing and TIL assessment on each image was performed.
  3. A training model was developed to generate TIL maps to ‘learn’ a classification (or predictive) model.
  4. The training model was then used to identify, classify and quantify new data elements to formulate higher-order relationships and clustering patterns within the entire sample set.
  5. TIL pattern results were compared with genomic assessments and linked with clinical outcomes for final analysis.

The computational pathology team at Reveal Biosciences is using deep learning to analyze whole slide pathology images. Highlights include TIL mapping in tumors, biomarker analysis of computationally segmented tissue structures (tumor/stroma, glomeruli, etc.) and correlating pathology data with clinical outcomes. This data can accelerate research and is also being used to develop computational diagnostics to benefit patients.