Integrating convolutional and graph neural networks for diagnosis and prognosis of Hematologic diseases

Background: Hematologic disease diagnosis, particularly for malignancies, heavily relies on labor-intensive and variable bone marrow smear analysis. Deep learning holds promise for accurate blood disease diagnosis. However, existing methods have challenges such as data dependencies and limited generalization. Moreover, research integration of bone marrow morphology with prognosis remains insufficient. Method: To overcome these limitations, this study proposes a novel approach that integrates convolutional neural networks (CNNs) and graph neural networks (GNNs). This combined framework facilitates comprehensive analysis of global features and spatial relationships among cells in bone marrow smears. Notably, the introduction of a learnable Transformer fusion layer enables multitask analysis for both diagnosis and survival prediction in malignant hematological diseases. Results: Validation results on public datasets demonstrate the efficacy of our approach, achieving a diagnostic AUC of 0.905±0.029 and a prognostic AUC of 0.780±0.047. These findings indicate successful mitigation of challenges in bone marrow smear analysis through CNNs and GNNs integration, potentially enhancing diagnostic accuracy and patient outcomes. Implications: The proposed method represents a significant advancement in hematologic disease diagnosis and prognosis analysis. This innovation holds promise for improving diagnostic accuracy, facilitating timely interventions, and ultimately enhancing patient outcomes. These findings contribute significantly to the evolving field of artificial intelligence in healthcare and pave the way for further enhancements in hematologic disease diagnosis and prognosis analysis.