Diffuse large B-cell lymphoma (DLBCL) remains the most common and heterogeneous type of non-Hodgkin lymphoma. Accurate diagnosis is crucial but intensive. AI has emerged in the field as a potential to support the digital pathology workflow. This study is a systematic review of the performance and clinical utility of AI applied to digital pathology and classification of DLBCL through articles published from 2020 to 2025. We followed the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 2020 guidelines. A total of 734 records were screened, and 11 studies met the inclusion criteria. QUADAS-2 and risk-of-bias VISualization(robvis) were used to assess bias. Data extraction included AI model architecture, diagnostic tasks, validation methods, and performance metrics. Eleven studies met the inclusion criteria, employing architectures such as convolutional neural networks, multiple instance learning, Vision Transformers, EfficientNet, U-Net, and HoVer-Net. Diagnostic metrics were consistently high: accuracies ranged from 87% to 100%, sensitivities from 90% to 100%, specificities from 52% to 100%, and area under the curve values up to 0.999. Several models outperformed pathologists in speed and precision, particularly in biomarker quantification and MYC rearrangement prediction. Risk of bias was low in index tests and reference standards, but patient selection was frequently rated as of high concern. AI-driven digital pathology demonstrates strong classification and diagnostic potential for DLBCL, achieving high accuracy across diverse methods and datasets. However, selection bias, limited external validation, and lack of standardization remain barriers. Continued research with multicenter, prospective validation is needed before routine clinical integration. Further research and standardization are needed for broader clinical integration in the field of digital pathology.
Luna et al. (2025) studied this question.