Background: In recent years there has been rapid growth in artificial intelligence (AI) applications for hematologic malignancies. We conducted a systematic review of clinical studies (2019–2025) evaluating AI tools in key focus areas: diagnosis via morphologic analysis, minimal residual disease (MRD) detection, acute myeloid leukemia (AML) subtyping, and laboratory workflow enhancement.

Methods: Relevant studies were identified and data extracted on tool name, primary function, dataset size, performance metrics, and clinical validation status.

Results: We identified multiple AI systems demonstrating high accuracy in classifying blood/bone marrow cells and detecting disease. For example, the Morphogo system (CNN-based bone marrow cell classifier) was trained on ~2.8 million cell images and achieved ~99% overall accuracy (sensitivity ~81%, specificity ~99.5%) in identifying >25 cell types. DeepHeme, an ensemble deep-learning model for bone marrow smears, was trained on ~50,000 cell images and matched or surpassed expert hematopathologists in accuracy while reducing slide review time from ~30 minutes to seconds. In AML subtyping, the self-supervised DinoBloom vision transformer model (trained on 380k cell images) attained a ~92% weighted F1-score for predicting AML subtypes, outperforming prior methods (~82%). For MRD detection, AI-assisted flow cytometry tools have reached expert-level performance. MAGIC-DR (machine-learning guided AML MRD analysis) integrated an XGBoost classifier with UMAP visualization and achieved a median AUC ~0.97 and 100% concordance with manual blast detection in a 25-sample test. In CLL, a deep neural network (DNN)–assisted MRD workflow showed 97.5% concordance with manual analysis (sensitivity 100%, specificity 95%), verified down to 0.002% MRD, and reduced analysis time by ~60%. Conclusions: AI tools for hematologic malignancies have demonstrated high diagnostic accuracy and substantial workflow improvements. Several systems (e.g. Morphogo, CLL MRD DNN) have undergone clinical validation and are being implemented in practice, whereas others (e.g. DinoBloom, MAGIC-DR, DeepHeme) remain in research or early validation stages. These advances highlight AI's potential to standardize hematopathology, enhance MRD monitoring, and assist in complex diagnostic tasks. Ongoing clinical validation and integration into lab workflows are critical to translate these AI tools into improved patient outcomes.

This content is only available as a PDF.
2025
Sign in via your Institution