A clinical AI-driven multiplex immunofluorescence imaging pipeline to characterize tumor microenvironment heterogeneity.

Abstract

3020 Background: Understanding the underlying heterogeneity of the tumor microenvironment (TME) on a single-cell level is becoming increasingly important to predict a patient’s response to immunotherapy. Conventional imaging methods can help reveal tissue heterogeneity, but are not optimal for identifying multiple cellular subpopulations or cellular interactions from a single slide image, limiting their use in clinical settings. Here, we present a clinical artificial intelligence (AI)-driven multiplex immunofluorescence (MxIF) imaging pipeline based on novel cell segmentation and cell typing methods to evaluate tumor cellular heterogeneity, immune cell composition, and cell-to-cell interactions. Methods: A machine learning (ML)-based cell segmentation algorithm was trained on a manually annotated dataset created from 219 different regions of interest (ROIs) that contained 85,991 cells from various tissues (colon, kidney, lung, lymph node, tonsil, and ureter). A dataset containing 58,676 cells from 146 ROIs was used for validation and accuracy was determined between automated and manually annotated images; accuracy was further evaluated by calculating the f1-score using available methods (DeepCell and Stardist). Marker stains with a low signal-to-noise ratio were automatically enhanced, allowing for adequate cell-to-cell interaction analysis. Results: An automated MxIF image processing workflow was developed. Validation of the trained cell segmentation model showed high accuracy (0.80 f1-score), demonstrating superior performance compared to other methods (DeepCell and Stardist - 0.55 and 0.78 f1-score, respectively). The pathologist-determined accuracy (0.84 mean f1-score) indicated a near-human performance of the developed method. Normalized expression values obtained from the cell typing model allowed automated cell recognition. We analyzed cellular heterogeneity across 3 regions of colorectal cancer (CRC), gastric cancer (GC), and non-small cell lung cancer (NSCLC) samples. While proportions of immune cells varied, proportions of malignant epithelial cells were stable across all regions of each sample, as concordant percentages of Ki67+ cells were identified (CRC-19%; GC-21%; NSCLC-5%). Analysis of cell-to-cell interactions and immune communities identified tumor-, immune-, and stromal-enriched communities in all tumor samples that were stable across regions. Conclusions: By analyzing complex tumor tissue at single-cell resolution with high accuracy, this AI-driven MxIF imaging technology is able to characterize tumor and microenvironment heterogeneity across cancer types. This novel AI-based tool is currently being integrated into several ongoing prospective clinical studies to aid in the development of predictive and prognostic biomarkers.