The time-consuming analysis of biomedical image data, e.g. in cell microscopy, is currently performed by subject matter experts and is accordingly personnel and cost intensive. Furthermore, the subjectivity of the evaluation and the susceptibility to application- and device-specific errors impair the comparability of the results. Automating this evaluation activity would allow biomedical experts, such as biologists, physicians, and virologists, to use their time more effectively for creative or communicative activities.
Deep learning (DL) methods, which use deep artificial neural networks for semantic knowledge extraction from image data, have been achieving impressive results in various domains and especially in the life sciences for several years as an extension of conventional image processing. However, developing such applications requires experience and expertise in data science, machine learning, information technology, and software development. Basically, specialists from outside the field rarely have the necessary skills to develop functional DL applications for their specific problems themselves. Moreover, commissioning specific individual solutions is often not economically justifiable for the institutions.
The goal of AIxCell is therefore to enable experts to train and use complete DL solutions for concrete problems themselves with the help of an AutoML software tool. The innovative character is formed by a domain-specific meta-learning system, which suggests a selection of best possible DL algorithms including data pre-processing, model selection as well as configuration and post-processing for the available resources and the specific use case the biologist or physician wants to solve. The selected model is then trained and made available to the expert for use. The core of the meta-learning system is a decision logic at the meta-level, called AutoKonfig, which outputs a selection of best-performing and most suitable algorithm configurations for the input data, use case, and resource requirements. In the course of the project, DL algorithms are developed to solve specific problems of the consortium partners and stored in a library together with the data sets, the requirements and the evaluation results. This library in turn forms the metadata set on which the decision logic is trained in a monitored manner. In the process, the decision logic learns to explicitly output a best possible pre-selection of algorithm configurations for a given task. Finally, the meta-learning system and the DL library are embedded in an application that provides the expert with a user-friendly front-end for entering and annotating the data set, specifying the task to be solved, displaying the results, and finally using the DL model.
- Fraunhofer Institute for Production Technology IPT
- ALS Automated Lab Solutions GmbH
- Bayer AG
- Cellmatiq GmbH
- Labforward GmbH
- MABRI.VISION GmbH
- MINDPEAK GmbH
- Olympus SIS GmbH
- ORACLE Germany B.V. & Co. KG
- PicoQuant GmbH
- Ruhr University Bochum
- Stem Cell Network NRW e.V.
- Taorad GmbH
- University Hospital Cologne
- University Hospital RWTH Aachen
- University Hospital Bonn
- University Medical Center Göttingen
For more information on the "AIxCell" project and other projects of the Research Association for Precision Mechanics, Optics and Medical Technology, please visit www.forschung-fom.de.