Automated Image Analysis Pipeline for Standardized Processing, Segmentation, and Feature Extraction in Diffuse Large B-Cell Lymphoma Histological Slides: Towards Enhanced Prediction of Immunotherapy Response and Risk Stratification

Introduction:

Diffuse Large B-Cell Lymphoma (DLBCL), the most prevalent form of non-Hodgkin lymphoma, presents significant treatment challenges due to heterogeneous responses and the need for rapid biomarkers to anticipate risk and treatment response, particularly to immunotherapy. Traditional histopathological methods often lack the precision required to understand tumor microenvironment interactions. Recent advancements in imaging techniques combined with artificial intelligence (AI) offer promising solutions for detailed analysis. These technologies could predict risk and identify predictive signatures of response to immunotherapy by measuring tumor and microenvironment interactions. This study introduces the LymphoPath platform, designed to enhance diagnostic and therapeutic approaches for DLBCL.

Objectives:

The primary objective of the LymphoPath platform is to harmonize image acquisition, processing, and analysis for DLBCL, improving diagnostic accuracy, understanding disease progression, and identifying potential biomarkers and therapeutic targets. The platform aims to provide a user-friendly, decentralized, and widely-available image repository to standardize DLBCL histology analysis across Europe.

Methodology:

We have developed an automated image analysis pipeline for standardizing image processing, segmentation, and feature extraction from DLBCL biopsy samples. The system begins by identifying nuclei, then classifies cells using a supervised model trained with 20,000 classifications. This is followed by feature extraction at both the single-cell and cell-cluster levels. Filters have been developed to automatically detect non-lymphoma tissue to prevent erroneous metrics from adjacent tissues and identify conflicting samples. This pipeline forms the foundation for a future collaborative, decentralized platform aimed at identifying digital biomarkers in DLBCL.

Results:

We have processed diagnostic images from 452 patients diagnosed between 2000 and 2021 and treated with R-CHOP or R-CHOP-like upfront regimens. Preliminary results demonstrate that our image analysis pipeline performs automated, detailed analyses of the tumor microenvironment, cellular morphology, and spatial relationships within the tissue. Our supervised classification model achieved 99% accuracy in classifying lymphoma and microenvironment cells. We have identified more than 200 distinct features to extract intensity, shape, and texture information from each cell individually and at the class level, as well as spatial relationships between tumor and microenvironment cells. These features include 4 shape statistics, 2 summary statistics, 2 cellular and tissue-level measurements, 2 distance metrics, and 4 Haralick texture features, extracted using the well-established tool QuPath. The feature extraction process has identified histological characteristics that correlate with treatment outcomes, aiding in the prediction of patient responses to immunotherapy and the risk of relapse. The unsupervised clustering approach has revealed distinct subtypes of DLBCL, each characterized by unique histological profiles.

Conclusion:

Our image analysis pipeline represents a significant advancement in the field of hematology, offering a comprehensive tool for the automated analysis of DLBCL. By standardizing image processing, segmentation, and feature extraction, this pipeline enhances the identification of key histological features and predictive biomarkers, ultimately improving patient outcomes. The future implementation of this pipeline as a collaborative, decentralized platform will further support the identification of digital biomarkers, standardizing DLBCL histology analysis, and enabling better-informed clinical decisions and personalized treatment strategies.