Application of Machine Learning for Automatic MRD Assessment in Paediatric Acute Myeloid Leukaemia

Roxane Licandro, Michael Reiter, Markus Diem, Michael Dworzak, Angela Schumich, Martin Kampel

Abstract: Acute Myeloid Leukaemia (AML) is a rare type of blood cancer in children. This disease originates from genetic alterations of hematopoetic progenitor cells, which are involved in the hematopoiesis process, and leads to the proliferation of undifferentiated (leukaemic) cells. Flow CytoMetry (FCM) measurements enable the assessment of the Minimal Residual Disease (MRD), a value which clinicians use as powerful predictor for treatment response and diagnostic tool for planning patients’ individual therapy. In this work we propose machine learning applications for the automatic MRD assessment in AML. Recent approaches focus on childhood Acute Lymphoblastic Leukaemia (ALL), more common in this population. We perform experiments regarding the performance of state-of-the-art algorithms and provide a novel GMM formulation to estimate leukaemic cell populations by learning background (non-cancer) populations only. Additionally, combination of backgrounds of different leukaemia types are evaluated regarding their ability to predict MRD in AML. The results suggest that background populations and combinations of these are suitable to assess MRD in AML.