Cristian Barrera
BME PhD Defense Presentation
Date: 2025-11-15
Time: 1:00 pm - 2:30 pm
Location / Meeting Link: https://zoom.us/j/91890336855?from=addon
Committee Members:
Anant Madabhushi, Ph.D. FAIMBE, FIEEE, FNAI (Advisor); Gari Clifford, DPhil, FIEEE,
Ahmet F. Coskun, Ph.D., Sunil S. Badve, MD, FRCPath, Kristin Higgins, MD
Title: Advanced Computational Pathology in Lung Cancer: From Immune Microenvironment Analysis to Predictive Biomarkers in Non-small and Small Cell Lung Cancer
Abstract:
Lung cancer remains a leading cause of cancer mortality, and both non-small cell (NSCLC) and small cell lung cancer (SCLC) require robust, scalable biomarkers that can be derived from routine histopathology to guide prognosis and therapy selection. This thesis develops and unifies two complementary classes of computational pathology models on hematoxylin and eosin (H&E) images: (i) “visible” pathomics signatures that explicitly quantify single-cell and spatial immune–tumor architecture, and (ii) “non-visible” AI signatures that learn latent immune activity, particularly CD8 cytotoxic T-cell biology, from high-dimensional image embeddings. PhenoTIL represents the first visible, TIL-centric niche signature, showing that immune cell clusters captured from H&E alone stratify NSCLC patients treated with chemotherapy and immunotherapy, and retain prognostic and predictive value when evaluated on clinical trial cohorts such as CheckMate 057. PhenopyCell extends this visible framework to whole-slide, region-aware spatial modeling in both NSCLC and SCLC, yielding interpretable features that independently associate with overall survival and treatment benefit, including platinum chemotherapy response in limited- and extensive-stage SCLC and immunotherapy benefit in NSCLC across multiple real-world cohorts and clinical trial datasets. In parallel, a non-visible CD8 predictor was engineered that learns patch-level CD8 activity directly from H&E images using immunofluorescence as biological ground truth and feature embeddings from histology foundation models and image-based gene expression predictors, enabling accurate estimation of spatial CD8+ T-cell signatures without additional staining; this CD8-derived signature was validated against CD8 immunofluorescence, associated with survival in multi-institutional NSCLC cohorts, and demonstrated treatment-selection utility within CheckMate 227, including PD-L1–high and PD-L1–low strata and immunotherapy versus chemotherapy arms. Together, the visible (PhenoTIL, PhenopyCell) and non-visible (CD8-based) models constitute an integrated, clinically evaluated toolbox that recovers immune microenvironment structure and function from standard histology in NSCLC and SCLC, supporting precise risk stratification and more personalized allocation of chemotherapy, immunotherapy, and their combinations in lung cancer.