Deep learning for leukemia diagnosis faces fragmented research and limited multimodal integration in current literatureAI for leukemia diagnosis still fragmented, review finds

The diagnosis of leukemia is inherently a multimodal process that combines morphological evaluation, immunophenotyping, molecular profiling and standard laboratory tests. Deep learning techniques have been adopted across all components of the diagnostic process over the past few years, particularly in microscopic image analysis. However, current research and review articles are highly fragmented, and most methods are limited to the single data modalities and benchmark-based performance analysis, which provide a limited understanding of model generalizations, clinical and real-world applicability. This scoping review presents a detailed analysis of recent studies of leukemia detection and diagnosis based on deep learning, in a structured and clinically grounded approach, and summarizes the literature on major diagnostic modalities, such as peripheral blood and bone marrow image analysis, gene expression and multi-omics modelling, flow cytometry-based learning, routine laboratory data analysis, and multimodal diagnostic approaches. However, only a limited number of studies have implemented true multimodal integration. Instead of just evaluating reported accuracy, this scoping review critically analyzes the impact of dataset design, modality-related bias, and validation methods, and model interpretability on the reliability and translational capabilities of the suggested approaches. New methodological developments, including transformer-based architectures, attention-based multi-instance learning, foundation models and explainable artificial intelligence, are presented in connection to robust, cross-dataset generalization and clinical reliability. Combining evidence representing traditionally separated lines of research, this scoping review highlights the research gaps in the methodology and the translation of AI methods to the real world, for the use of AI in leukemia. Moreover, it provides a clear roadmap for research in this area, emphasising modality-aware fusion, external validation, regulatory-aware, and human-in-the-loop decision support.