1276 Leveraging artificial intelligence (AI) models delineating tumor vs immune cell expression for scalable biomarker analysis of clinical trial samples: a digital image analysis approach for NSCLC

BackgroundStandard, manual, single-pathologist histopathology assessment in clinical trials makes accurate quantification of biomarkers expressed by both tumor and immune cells challenging due to inter-pathologist variability and non-exhaustive cell evaluation. Artificial intelligence (AI) models offer quantitative, reproducible solutions in optimized patient characterization for precision therapeutics but typically require large sample sizes for development. Here, we investigated whether a digital image analysis pipeline could be developed to analyze non-small cell lung cancer (NSCLC) from a first-in-human (FIH) Phase I clinical trial with limited samples, variable tissue content, and pre-analytical parameters addressing the outlined challenges for (1) H&E-based tumor microenvironment (TME) characterization and (2) scoring aryl hydrocarbon receptor (AhR) expression in serial immunohistochemistry (IHC)-stained sections.MethodsThe study used 25 H&E and 24 IHC (AhR) NSCLC biopsies (NCT04069026, following standard ethical guidelines). Pre-trained NSCLC models for H&E-based tissue segmentation (carcinoma, stroma, necrosis, ‘other’) and cell detection and classification (carcinoma vs ‘other’ cells) were optimized on 20 and evaluated on five hold-out slides. AI models were developed for IHC-based cell detection and classification (carcinoma vs ‘other’) and nuclear AhR expression prediction (negative, weak, moderate, strong). Pathologists provided training annotations using detailed guidelines. Evaluation annotations in pre-defined regions were withheld from training. AhR scores and spatial statistics were derived via automated TME analysis. Pathologists were assessed for inter-rater variability in IHC-based cell classification and AhR scoring tasks.ResultsThe H&E tissue segmentation and cell classification models demonstrated mean F1 scores of 0.92 and 0.83, respectively. The IHC-based cell classification model achieved an F1 score of 0.90. The AhR expression model exhibited an F1 score of 0.93 for ‘AhR present vs not present’ and a mean F1 score of 0.80 across all intensity categories. Inter-rater agreement among three pathologists was moderate for carcinoma classification (Krippendorff’s alpha [KA] 0.77). Moderate agreement for AhR nucleus negativity (KA 0.77) and weaker agreements for the other categories were observed (KA weak expression 0.42, moderate expression 0.32, strong expression 0.69).ConclusionsInter-rater variability results underscore the range of