You are here: Home Press releases Breakthrough in AI based Flow Cytometry
Date: Apr 06, 2020

Breakthrough in AI based Flow Cytometry We were able to achieve expert level accuracy in solving a multiclass flow cytometry problem.This was possible by transforming the problem into self organizing maps allowing us to train a deep convolutional neural network to perform the task. The software of our classifier is open source and we make our entire model publicly accessible via https://github.com/xiamaz/flowCat. A ready-to-use webservice is available at https://hema.to.

In Hematologist-level classification of mature B-cell neoplasm using deep learning on multiparameter flow-cytometry data we demonstrate the first application of artificial intelligence that solves a multiclass problem in flow cytometry with expert accuracy. We were able to do so, because we applied advanced machine learning technology on a big data set, that is samples from the Munich Leukemia Laboratory comprising 20622 patient samples. On multiparameter flow-cytometry (MFC) data from Navios flow cytometers, we achieved an F1 score of 0.94 on seven B-cell non Hodgkin lymphoma (B-NHL) subtypes and healthy individuals. Our multiclass classification problem includes the diagnoses of Chronic Lymphocytic leukemia and Monoclonal B-cell lymphocytosis (CLL/MBL), Marginal zone lymphoma (MZL), Mantle cell lymphoma (MCL), Prolymphocytic leukemia (PL), Follicular lymphoma (FL), Hairy cell leukemia (HCL) and Lymphoplasmocytic lymphoma (LPL).

The key challenge that we had to solve, was making the multiparameter flow-cytometry data amenable for a deep convolutional neural network (DCNN). We did so by transforming MFC data into a self-organizing map, that allows an image analysis in 2D. By this means we redefined the problem so that pattern recognition by DCNNs can be performed.

 The results of our work show that current systems of artificial intelligence can be used to increase the quality and efficiency in a routine setting of clinical flow cytometry. Furthermore, even in inconclusive cases, the signals picked up by our network, might be important decision support for the human expert. The software of our classifier is open source and we make our entire model publicly accessible via https://github.com/xiamaz/flowCat. A ready-to-use webservice is available at https://hema.to.

Document Actions