Clinical implementation readiness of machine learning in hematologic malignancy imaging: A scoping review and evidence-based framework

Abstract

Abstract Background Machine learning (ML) tools for hematologic malignancy imaging show strong technical performance, yet remarkably few have reached clinical use. Given rapid growth in this field, we sought to systematically assess current implementation readiness and identify barriers to translation. Methods We conducted a scoping review of studies applying ML to radiologic imaging of hematologic malignancies (CT, MRI, PET/CT, X-ray). Eligible studies reported performance metrics and addressed at least one readiness factor: validation method, dataset characteristics, workflow considerations, or regulatory preparedness. We extracted study design, ML approach, validation strategy, imaging modality, and readiness indicators, classifying each as technical, preclinical, early clinical, or workflow pilot. Results Our analysis included 25 studies spanning multiple hematologic malignancies and imaging approaches. While technical validation was universal, translation to clinical use lagged sharply:Clinical validation: 12/25 (48%)Rigorous reader studies or real-world testing: 4/25 (16%)Workflow integration pilots: 1/25 (4%)Regulatory clearance: 0/25 (0%)External validation on independent cohorts: 5/25 (20%)Radiologist/clinician comparison: 4/25 (16%) Perhaps most striking, only a single study piloted workflow integration, and none reported regulatory clearance. The included studies represented substantial development efforts, with a median dataset size of 650 patients (range ~100–2400) and showed growing momentum through ~25% annual publication growth between 2020–2024. Deep learning approaches were most common (n=9), followed by radiomics (n=8) and support vector machines (n=4). Multiple myeloma dominated disease representation (n=7), with lymphoma also prominent (n=3). CT (n=6), MRI (n=5), and PET/CT (n=6) were equally represented as imaging modalities. Key barriers included absence of regulatory preparedness, rare workflow integration, lack of standardized imaging and reporting, limited prospective multicenter validation, and insufficient model explainability. Conclusions Despite universal technical validation, ML in hematologic imaging remains predominantly preclinical. The stark contrast between strong technical performance and near-absent workflow integration (4%) or regulatory engagement (0%) highlights urgent translation gaps. Recommendations Prioritize multicenter validation for high-performing tools (>90% accuracy) to confirm generalizability. Pursue early FDA consultation and Breakthrough Device designation for advanced applications (≥90% accuracy with external validation). Standardize imaging protocols and reporting to enable reproducible validation. Integrate clinician-in-the-loop development with explainable outputs for workflow adoption.