ARTIFICIAL INTELLIGENCE IN BREAST CANCER IMMUNE INFILTRATION PATHOLOGY IMAGING

Breast cancer prognosis and treatment response depend on tumor-infiltrating lymphocytes (TILs) within the tumor microenvironment, but manual assessment of H&E-stained slides suffers from inter-observer variability and labor intensity. This review paper synthesizes current evidence on artificial intelligence (AI), particularly deep learning, for automated TIL quantification in whole slide images (WSIs) of breast cancer pathology. The review examines convolutional neural networks (e.g., U-Net, StarDist) for cell detection and segmentation, foundation models (e.g., ECTIL) for label-efficient learning, open-source pipelines (e.g., QuPath), and multimodal approaches integrating spatial TIL patterns, tertiary lymphoid structures, and multiomics data. AI methods show strong concordance with expert pathologist assessments across multi-center cohorts, provide independent prognostic value for patient survival, and offer improved prediction of neoadjuvant chemotherapy response over manual scoring. Label-efficient foundation models reduce annotation demands while preserving clinical utility, and spatial analyses of TIL distribution and cellular interactions yield additional biological insights. Challenges encompass staining variability, generalization across breast cancer subtypes and populations, and clinical workflow integration. Future directions include standardization efforts, regulatory approval pathways, real-time decision support for pathologists, and broader application to multiplexed immune profiling for immunotherapy guidance. This review highlights AI's potential to standardize TIL assessment and advance precision breast cancer pathology.

Перевод пока недоступен