Immunotherapy Prediction Challenge Immune checkpoint inhibitors targeting PD-1 and PD-L1 have transformed lung cancer treatment, but only a subset of patients responds. PD-L1 protein expression, measured by immunohistochemistry (IHC), is the current standard biomarker for patient selection, but IHC requires additional staining beyond the routine hematoxylin and eosin (H&E) stain used in all diagnostic pathology.
Pathomics Approach This study by Dia and colleagues explored whether computational analysis of standard H&E whole slide images (WSIs) - already collected for every patient - could predict both PD-L1 expression level and actual treatment outcome (progression-free survival) without needing additional IHC testing. This approach, called pathomics, extracts quantitative texture and structural features from tissue images.
Clinical Value If H&E-based computational analysis can accurately predict immunotherapy response, it could eliminate the need for separate PD-L1 IHC testing, reduce costs and turnaround time, and enable prediction even from archival tissue where fresh IHC cannot be performed.
Haralick Feature Extraction The researchers extracted Haralick texture features from H&E WSIs. Haralick features mathematically describe how pixel intensities are spatially related to each other in an image - capturing patterns of tissue architecture, cellular density, and morphological regularity that the human eye processes intuitively but AI can quantify precisely.
Feature Selection Multiple feature selection algorithms (SurfStar, MultiSurf, and others) were applied to identify which of the many extracted Haralick features were most predictive of PD-L1 expression and PFS. Feature selection reduces dimensionality and prevents overfitting to noise in the training data.
Cohort Design Training was performed on 43 patients from CHUM (Centre Hospitalier de l'Universite de Montreal) and validation on 25 independent patients from IUCPQ (Institut Universitaire de Cardiologie et de Pneumologie de Quebec), providing a proper external validation split.
PD-L1 Prediction The best model for PD-L1 expression prediction used a Support Vector Machine (SVM) classifier with SurfStar feature selection, achieving AUC of 0.66 in validation. While modest, this is statistically significant and demonstrates that H&E texture alone carries some information about underlying protein expression.
PFS Prediction For predicting progression-free survival, a Linear Discriminant Analysis (LDA) classifier with MultiSurf feature selection performed best, achieving AUC of 0.71 in the training cohort and 0.62 in validation. Predicting actual treatment outcomes is a harder task than predicting a protein expression level.
Biological Validation The researchers used Gene Set Enrichment Analysis (GSEA) to validate their findings at a molecular level. Genes associated with the high-pathomics-score tissue regions were enriched for NK cell cytotoxicity, chemokine signaling, and T-cell receptor signaling pathways - exactly the immune pathways expected to be active in regions with high PD-L1 and favorable immunotherapy response.
Tissue Microenvironment The immune microenvironment of a tumor - whether it is inflamed with dense lymphocyte infiltration or cold and excluded - produces visible architectural changes in the tissue that H&E staining captures. More lymphocytes, altered glandular structures, and stromal remodeling all change the texture statistics that Haralick features measure.
Connection to PD-L1 Tumors that express high PD-L1 often do so in response to interferon-gamma from infiltrating T cells - a process called adaptive immune resistance. This immune infiltration alters tissue architecture in ways that computational texture analysis can detect, even without explicitly visualizing the PD-L1 protein.
Practical Implication The correlation between Haralick features and immune pathway activity (confirmed by GSEA) validates that pathomics is not simply pattern-matching to spurious correlations, but is genuinely detecting biologically relevant tissue organization.
Sample Size Constraint The 68-patient cohort is small for a machine learning study, limiting statistical power and the ability to capture rare but important subgroups. Larger multi-institutional datasets are needed to more reliably estimate predictive accuracy.
Complementary Role Rather than replacing PD-L1 IHC, pathomics analysis could complement it - for example, when IHC results are borderline or when tissue is insufficient for additional staining. It could also provide additional independent prognostic information beyond the binary PD-L1 score.
Future Development Incorporating deep learning-based feature extraction (convolutional neural networks trained end-to-end on WSIs) alongside traditional Haralick features could capture higher-order spatial patterns and potentially improve prediction accuracy beyond the current AUC values.