AI’s deep learning algorithms analyze medical samples with automation and standardization

Researchers from the Helmholtz Zentrum München in Munich and the University Hospital of LMU Munich have shown for the first time that a deep learning algorithm performs similarly to human experts when classifying blood samples from patients with acute myeloid leukemia (AML). Their proof-of-concept study paves the way for automation, standardization and field sample analysis in the near future. The paper was published in Nature Machine Intelligence.


AI’s deep learning algorithms analyze samples in an automated and standardized way. Left: Classification of human experts. Right: Pixels important for AI analysis.

Every day, millions of single cells are diagnosed in medical laboratories and clinics for disease diagnosis. Most of the repetitive tasks are still done manually by trained cytologists, who examine cells in stained smears and classify them into about 15 different categories. There is taxonomic variability in this process and requires the presence and expertise of a trained cytoologist.

To improve evaluation efficiency, a team of researchers at Helmholtz Zentrum München and the LMU University Hospital in Munich trained a deep network of neurons with about 20.000 single-cell images to classify them. The team, led by Dr. Carsten Marr and MD student Dr. Christian Matek from the Institute for Computational Biology at Helmholtz Zentrum München, and Prof. med Karsten Spiekermann and Simone Schwarz from the Third Department of the LMU University Hospital Munich, the images were obtained from 100 patients with invasive Blood smears were extracted from patients with hematologic disease AML and 100 controls. The new AI-driven approach is then evaluated by comparing its performance with the accuracy of human experts.

Deep learning algorithms for image processing require two things: first, a proper convolutional neural network architecture with hundreds of thousands of parameters; and second, a sufficiently large amount of training data. To date, there are no large digitized blood smear datasets, despite the common clinical use of these samples. The research group at Helmholtz Zentrum München has now provided the first large dataset of this type. Currently, Marr and his team are working closely with the Third Department of Medicine at the LMU Munich University Hospital and with the Munich Leukemia Laboratory (MLL), one of the largest European leukemia laboratories, to digitize hundreds of patient blood smears.

To bring our method to the clinic, digitization of patient blood samples has become routine. Algorithms must be trained using samples from different sources to cope with the inherent heterogeneity in sample preparation and staining. Together with our partners, we can demonstrate that deep learning algorithms show similar performance to human cytoologists. As a next step, we will evaluate how other disease characteristics, such as genetic mutations or translocations, can be predicted using this new AI-driven approach. “

The method demonstrates the application capabilities of AI in translation studies. This is an extension of Helmholtz Zentrum München’s pioneering work on single-cell classification of blood stem cells (Buggenthin et al., Nature Methods, 2017), which was awarded the Erwin Schroedinger Prize from the Helmholtz Society in 2018. Supported by SFB 1243 of the German Research Foundation (DFG) and a PhD scholarship to Dr. Christian Metek from the Jose Carreras Leukemia Foundation of Germany.

The Links:   NL6448BC26-01F G185XW01-V201