Periodic Reporting for period 2 - CompHematoPathology (Computational Hematopathology for Improved Diagnostics)
Période du rapport: 2021-12-01 au 2023-05-31
Detecting serious blood diseases in microscopic images of individual cells requires a lot of experience, patience and hard work: Every day, cytologists around the world examine and classify hundreds of individual blood cells based on their morphological features to diagnose a blood sample. This is obviously a prototypical challenge for modern Deep Learning-based image analysis methods. Automating single cell classification, which suffers from human error and expert variability, would standardize the process and free up missing expert capacity in our strained healthcare system.
With the help of the ERC CoG grant, I am addressing this problem. Together with my team, we are designing and training machine learning models that are able to robustly classify individual cells and thus contribute to the diagnosis of serious blood diseases such as leukemia. Furthermore, we aim to better understand the basis for these diseases and are thus developing mathematical models that can reproduce the kinetics of blood production.
With the help of the ERC CoG grant, I am addressing this problem. Together with my team, we are designing and training machine learning models that are able to robustly classify individual cells and thus contribute to the diagnosis of serious blood diseases such as leukemia. Furthermore, we aim to better understand the basis for these diseases and are thus developing mathematical models that can reproduce the kinetics of blood production.
Besides microscopic blood smears, the analysis of bone marrow smears in blood cancer patients is a central part of the diagnostic pipeline. Together with partners at the Fraunhofer Institute in Erlangen and the Munich Leukemia Laboratory MLL in Munich, we were able to show that a machine learning model trained with over hundred thousand single cell images is able to recognize cell types as well as a human expert.
We use a similar model for the analysis of pathological tissue sections, whose sheer size and complexity (one section shows millions of individual cells) challenges pathologists on a daily basis. Applied to the analysis of subtypes of lymphoma, a group of blood and lymph tumors, we have been able to generate promising initial results, which we are currently validating on a second data set from a different hospital.
To gain insights into changes in the production of blood cells in human bone marrow, we have fitted mathematical models to experimental data from our biomedical collaborators. This approach allows us to identify specific differentiation patterns in time-resolved single cell data.
We use a similar model for the analysis of pathological tissue sections, whose sheer size and complexity (one section shows millions of individual cells) challenges pathologists on a daily basis. Applied to the analysis of subtypes of lymphoma, a group of blood and lymph tumors, we have been able to generate promising initial results, which we are currently validating on a second data set from a different hospital.
To gain insights into changes in the production of blood cells in human bone marrow, we have fitted mathematical models to experimental data from our biomedical collaborators. This approach allows us to identify specific differentiation patterns in time-resolved single cell data.
The application of machine learning models to clinical data requires special algorithms. We develop these and adapt them to the requirements of the clinical question and the available data sets. In the process, new methods, new questions, and new opportunities arise. In the further course of the project, we want to deepen our understanding of hematopoiesis, improve the diagnosis of hematopoiesis, and contribute to new forms of therapy for severe diseases.