Comprehensive bone marrow analysis integrating deep learning-based pattern discovery (BMDeep)

Bone marrow morphology forms the basis for the assessment of hematopoiesis. The currently established approach is dependent on manual microscopic counting of a limited number of cells by specially trained personnel, which is inherently associated with substantial intra- and inter-individual variability. The aim of our project is to automatize and improve the evaluation of bone marrow smears and to identify pathological patterns in pediatric leukemia.

Bone marrow smears from acute lymphoblastic, acute myeloid and chronic myeloid leukemia were scanned with high resolution after identification of hard- and software requirements for the scanning process by evaluating ten different scanner models. We developed an annotation tool for large scale classification of hematopoietic cells. We then designed a supervised Deep-learning model that is currently trained using the leukemia dataset. Next steps are the implementation of clinical data to enhance the capabilities of our model and allow for multi-dimensional pattern recognition. These patterns might identify new biomarkers to further enhance our understanding of pediatric leukemia.

Publication History

Article published online:
17 May 2022