Histogenomic analysis of tumor immunologic phenotypes

Customer: Boehringer Ingelheim (Ingelheim am Rhein, Germany)

Summary:  The project aims to classify the immunological phenotype of human solid tumors, which is essential for predicting disease progression and assessing response to immunotherapy. These tumors are typically classified into three immunological phenotypes: immune inflamed, immune excluded, or immune desert. To achieve this, we developed an end-to-end (E2E) pipeline for the automated analysis of histology images, designed to identify and classify stromal, lymphocyte, and cancer cells. The workflow comprises two main stages: cell segmentation and cell classification, leveraging a combination of machine-learning models and expert annotations. The initial dataset, annotated by our team of histopathologists, provided a solid foundation for further analysis.

The proposed pipeline was trained and evaluated on a comprehensive dataset of 76 whole slide images (WSIs), which were segmented into 527 patches containing 7318 nuclei. We evaluated existing cell segmentation tools such as cGAN and YOLOv8 to benchmark our pipeline's performance. The cell segmentation stage employed advanced machine learning techniques, HoVer-Net, to accurately delineate cancer cells, while the AutoML-based classification stage achieved a weighted F1-score of 79.8% in distinguishing between various cell types.

Additionally, the pipeline includes a tumor-immune classification module that categorizes WSIs into immune phenotypes (desert, excluded, inflamed, or unknown) with high accuracy and reproducibility. This categorization has significant potential for clinical application, aiding in the development of personalized treatment strategies.

The proposed E2E pipeline for histopathology image analysis demonstrates high accuracy, efficiency, and potential for reducing operational costs. It provides a robust tool for medical laboratories, enabling quick and reliable analysis of tissue samples, and supporting the diagnosis and treatment of cancer.

Collaborators: Anton Makoveev (Quantori, Prague, Czech Republic), Maksim Kazanskii (Quantori, Tbilisi, Georgia)  Di Feng (Boehringer Ingelheim, Ingelheim am Rhein, Germany)

Project type: Commercial

Media: Blog post

Figure 1. Schematic representation of the proposed solution for classifying the immunological phenotype of human solid tumors.