George Lee, PhD
Digital Pathology Informatics Lead
Bristol Myers Squibb
Machine learning classification of CD8-topology on 4000 tumor biopsies and resections via AI-powered image analysis of CD8 immunohistochemistry slides
Background: CD8-topology is an increasingly relevant area of research shown to stratify patient outcomes in solid tumors based on spatial CD8+ cell patterns. Understanding the role of CD8-topology in different clinical settings may present more personalized treatment options for patients. Conducting such studies in a reproducible way is challenging because manual interpretation of these complex patterns is subject to significant inter-reviewer variability. However, machine learning (AI) can quantify CD8-topology in a biologically-meaningful, reproducible, and scalable way. We have demonstrated the use of such AI methodologies to assess CD8-topology in 4,162 clinical and commercial CD8 IHC slides for melanoma (MEL), head and neck squamous cell carcinoma (HNSCC), and urothelial carcinoma (UC). Methods: We trained random forest AI-classifiers to predict pathologist-assigned inflamed, excluded, and cold patterns on CD8-IHC slides using parenchymal and stromal CD8 measurements from PathAIâ€™s deep learning platform. For validation, multiple pathologists scored CD8-topology in an independent set of 140 images, and we compared pathologist-pathologist concordance with pathologist-AI concordance. Results: Data from the validation set showed a range of inter-pathologist concordances measured by Cohenâ€™s kappa to be k=0.65 for MEL, k=0.86 for HNSCC, and k=0.57 for UC. The AI model performed similarly to pathologists, showing k=0.79 for MEL, k=0.66 for HNSCC, and k= 0.49 for UC. Conclusion: These results suggest that AI can accurately assess CD8-topology on multiple tumor types while avoiding inter-pathologist variation from manual scoring. Future work aims to leverage this capability to more efficiently study CD8-topology and its role in treatment outcomes and mechanisms of action.
- Understand the concept of immune infiltration as it relates to its appearance on histology slides
- Appreciate the difficulty of manual interpretation of immune infiltration patterns
- Understand the value of an automated approach to tissue interpretation
George Lee holds a PhD in Biomedical Engineering from Rutgers University and has authored over 50 peer-reviewed papers and patents around the topics of digital pathology image analysis, machine learning, and computer aided diagnosis. He previously served as a Research Assistant Professor at Case Western Reserve University (CWRU) in the Department of Biomedical Engineering. While at CWRU, he worked with clinicians and computer scientists to develop machine learning and image analysis algorithms to predict disease outcomes in prostate cancer from H&E-stained slides. At Bristol Myers Squibb, George collaborates broadly to mine pathology images for biologically relevant insights and developing image-based biomarkers which can identify patients who will benefit from immunotherapy treatments.