PD-L1 positivity in tumor-associated macrophages has been related to a favorable response to anti-PD-1 and PD-L1 targeted therapies. This concept has driven interest in specifically identifying macrophages in PD-L1 stained tissues, but pathologists often have difficulty in performing this task reliably. Thus, there is a clear need for tools capable of detecting and classifying tumor-associated macrophages in the context of a standardized PDL1 assay. One such method relies on the utilization of image analysis (IA) which measures hundreds of defining cellular features. This is necessary because macrophages identification based on few features would be difficult due to their varying morphologies and similarity with tumor cells. In order to isolate macrophages using multiple aspects of their cellular characteristics, AI and machine learning algorithms were leveraged to provide robust models for their identification. In this study we compare two artificial intelligence (AI) assisted macrophage classification methods for their accuracy in specifically identifying CD68 positive macrophages. Both approaches are based on interrogating a cohort of 23 non-small cell lung cancer (NSCLC) sections with two tumor-associated macrophage recognition algorithms, based on a decision tree (DT) model. These classifiers were developed by training a learning model on the morphometric, spatial, and chromogenic PD-L1 staining features measured on detected cells, using their fluorescent CD68 expression profile as an indicator of their macrophage lineage. These classification algorithms were trained on images developed using procedures wherein tissue sections were stained by immunofluorescence with antibody to the macrophage associated CD68 marker and subsequently scanned to obtain a digital image. After removal of the coverslip and stripping of the CD68 antibodies, the sections were then stained with the anti-PDL1 SP263 antibody and rescanned to obtain a second digital image. Co-registration of the two digital images allowed for the identification of CD68 positive cells in the context of only the SP263 PDL1 assay. The accuracy of each classification algorithm was determined by comparing the positivity for macrophage lineage among algorithm predicted macrophage to CD68 positivity. For the decision tree classifier, we demonstrated a > 90% accuracy classification of macrophage identification. Application of the SP263 model to NSCLC tissues stained with the 22C3 PD-L1 assay demonstrates the need for assay-specific detection models. The decision tree method achieved highly accurate predictions. These data indicate that sophisticated assay and data science methods allow for independent identification of macrophages using only a SP263 immunohistochemical PD-L1 assay, making it possible to isolate the contribution of PD-L1 macrophage positivity to successful checkpoint therapy response.
RESULTS & CONCLUSIONS
- An average 96.83% accurate model for detecting and quantifying macrophages (CD68-positive cells) in NSCLC tissues stained for PD-L1 (SP263) was achieved, improving on a top manual accuracy of 25%.
- Per-cell data profiles achieved via proprietary image analysis can be combined with thoughtful biological approaches and assays to remove manual bias for training cell detection models.
- Models built using SP263 staining are non-transferrable to 22C3-stained tissues, meaning models must consider assay and indication when used in practice.
- Application of the macrophage detection model in PD-L1 stained tissues allows investigators to interrogate PD-L1 status of macrophages without the need for additional stains. Pathologist review of markup accuracy is challenging based on difficulty of manual evaluations
- Combinatorial assay design and data science approaches to image analysis modeling will be expanded to multiple assays and tissues in future studies