AI in lymphoma diagnosis: high hopes and hard truths on the path to clinical reality

Diagnostic hematopathology is an attractive domain for artificial intelligence (AI) because lymphoid neoplasms are common, morphologically diverse, and increasingly defined by complex molecular features. Over the past decade, deep learning models have achieved impressive performance in classifying lymphoma from whole-slide images approximating immunohistochemical and genetic alterations from hematoxylin–eosin sections and extracting prognostic information that may be invisible to human observers. Yet despite this rapidly expanding literature, essentially no lymphoma-specific AI system is embedded in routine diagnostic workflows. This narrative review synthesizes recent advances in AI for lymphoma, with a primary focus on digital histopathology. We first outline key algorithmic paradigms that underpin modern computational pathology, including multiple instance learning, convolutional neural networks, vision transformers, and approaches to explainability and uncertainty estimation. We then summarize disease-specific applications across diffuse large B-cell lymphoma, other B-cell lymphomas, Hodgkin lymphoma, and T–/NK-cell lymphomas, highlighting both diagnostic and prognostic uses. Finally, we examine the main barriers that prevent translation of promising models into clinical reality, such as domain shift, label noise, limited external validation, workflow misalignment, regulatory and reimbursement uncertainty, and global inequities in data and infrastructure. On this basis, we propose a pragmatic implementation framework in which AI acts as a “digital resident” that supports triage, guided review, and quantitative augmentation rather than attempting full automation of lymphoma diagnosis. The adoption of AI in lymphoma diagnosis will depend less on new model architectures than on rigorous validation, thoughtful workflow integration, and ethically grounded, globally inclusive deployment strategies.