On automated Flow Cytometric analysis for MRD estimation of Acute Lymphoblastic Leukaemia: A comparison among different approaches

Minimal Residual Disease (MRD) is a powerful risk-stratification marker in the treatment of primary and relapsed childhood Acute Lymphoblastic Leukemia. Flow Cytometry is a fast and sensitive method to detect MRD, however, the interpretation of these multi-parametric data demands intensive operator training and experience. In this paper we propose a comparison among three different approaches for the automatic detection of leukemic cells in FCM data samples. The first approach is a fully-discriminative baseline based on the Support Vector Machine. The second approach is divided in two phases, the first is an unsupervised feature learning using a Stacked Auto-Encoder Neural Network, the second phase consists of exploiting the previously trained Neural Network to perform inference on the data classification. The third approach is fully generative and based on the Gaussian Mixture Model estimation and Bayes decision. Results show that a generative model provides better accuracy on MRD estimation with respect to the other two approaches. All samples are collected by national diagnostic reference centers for children with ALL for medical purposes enrolled in the current treatment protocol AIEOP-BFM-2009.