Artificial intelligence applications in digital onco-hematology

Artificial intelligence and digital pathology have revolutionized histopathology by converting images into high-resolution digital formats and introducing computational methods for image analysis. This thesis contributed to this context by aiming to stratify patients with various hematological conditions based on bone marrow fibrosis. The study involved the quantification of the percentage of bone marrow fibrosis on the hematopoietic tissue, used as indicator for characterizing different hematological disorders on a dataset of 1123 whole slide images of bone marrow biopsies. Hematopoietic tissue has been segmented through a color decomposition technique, while the segmentation of the fibrosis needed a more powerful segmentation model, which was trained on an active semi-supervised learning (ASSL) pipeline. Thanks to the ASSL approach it was possible to overcome the lack of annotated patches, resulting in more effective segmentation. Masks obtained from segmentation were used to calculate fibrosis and tissue areas. Final comparative analysis between control and disease revealed lower fibrosis percentages in the control group. With respect to control, Acute Myeloid Leukemia and Myelodysplastic Syndrome showed no significant differences, while Myeloproliferative Neoplasms exhibited significant differentiation, underscoring the role of bone marrow fibrosis in these conditions.