FDLP - Federated Digital Lymphoma Pathology

Aims of this collaborative project

Lymphoma is a heterogenous cancer entity composed of multiple morphological and molecular subtypes. To determine the optimal treatment, these subtypes need to be distinguished/diagnosed, which is a challenging task for pathologists. The state of the art in lymphoma diagnosis is to combine visual inspection of tissue with assessment of molecular data and merging these different types of data to identify the exact diagnosis. We and others have built two of the largest lymphoma data collections worldwide: one for molecular (OMICS) data and one for high resolution images. With these resources, we contributed to shaping the classification of lymphomas which currently guides study design and treatment decisions. As a next step, we want to bring both data sources closer together. A big obstacle on this route is the lack of computational approaches that allow us to automatize the joint modelling of these data. Moreover, state of the art histology image analysis by deep learning requires huge data collections for training. Luckily, suitable data are increasingly generated by pathologists across Germany during their daily practice in terms of scanned microscopic slides. However, data files are large and as health data obviously underlie safety regulations complicating their exchange. Thus, our primary goal is to establish a federated learning platform that integrates high-throughput molecular and high-definition histology image data. Sensitive data will remain with the local pathologists and do not have to be send to other sites; only model parameters will be shared. We expect the network to grow, as local pathologists increasingly use digital image data rather than physical tissue sections as a consequence of the COVID19 pandemic and the increasing use of home-office work among pathologists. Model stability will become a key challenge that we address by harmonizing image data from new providers in parameter space without exchanging original data. For more information, please see: BMBF FDLP

Overview of the Work Packages:

WP1: Computational Infrastructure

WP2: Machine Learning Methodology

WP3: Detection of Double Hit Lymphomas

Contribution of the Department of Medical Bioinformatics

Prof. Altenbuchinger  and his group coordinates the Workpage 2: Machine Learning Methodology
This comprises extraction of clinically relevant features from the trained in the FAI models (WP5), statistical evaluation and interpretation of those features. 
We will apply and extend in-house developed methods to interpret machine learning models especially in the context of molecular networks. Here these approaches will
elucidate potential causes and molecular mechanisms that drive the predictive performance of classifiers in PDAC

Project Partners