Multimodal AI in Tissue Diagnostics: Vision-Language Models and the Future of Computational Pathology.

Abstract

Vision-language models (VLMs) represent an emerging class of multimodal artificial intelligence (AI) systems that integrate visual information with natural-language understanding and generation. In computational pathology, VLMs provide a framework for aligning histologic morphology from whole slide images (WSIs) with pathology reports, and other text-based knowledge sources. This review summarizes the technical foundations, major applications, evaluation strategies, and deployment considerations of pathology VLMs. Current pathology VLMs support a growing range of use cases, including image-text retrieval, label-efficient classification, visual question answering, abnormality localization, anomaly detection, report generation, and agentic workflow support. These capabilities are enabled by image encoders, text encoders or large language models, multimodal alignment strategies, and, in some systems, generative language components. Despite rapid progress, several barriers remain. Evaluation of pathology VLMs is constrained by limited domain-specific benchmarks, insufficient assessment of visual grounding, overreliance on text-based metrics, vulnerability to hallucination, and uncertain robustness under data shift. Clinical translation also requires validation across institutions, scanners, staining protocols, tissue types, and patient populations, together with workflow integration, regulatory oversight, data privacy, cybersecurity, and pathologist accountability. VLMs are therefore best viewed as assistive systems that may augment rather than replace pathologists. Responsible development will require close collaboration among pathologists, computational scientists, health systems, and regulatory stakeholders to ensure that VLMs improves pathology practice in a safe, interpretable, and clinically meaningful manner.