Skip to content. | Skip to navigation

Sections
Personal tools
You are here: Home Theses A Meta learning Approach for Automating Image Analysis in Cell Biology

Contact

Prof. Dr. S. Decker
RWTH Aachen
Informatik 5
Ahornstr. 55
D-52056 Aachen
Tel +49/241/8021501
Fax +49/241/8022321

How to find us

Annual Reports

Disclaimer

Webmaster

 

 

A Meta learning Approach for Automating Image Analysis in Cell Biology

Thesis type
  • Master
Status Running
Submitted in 2021
Supervisor(s)
Advisor(s)

Biomedical image analysis in areas such as cell biology is an interdisciplinary study and has critical applications for healthcare investigation, life sciences analytics, and modeling medical challenges. These analyses are mostly done based on the manual assistance of domain experts based on their expertise and former theoretical concepts. Such manual efforts are not only time-consuming and have high chances of potential mistakes. However, the potential of AI lies primarily in the automation of manual activities and complex decision-making processes, and the capability of extracting knowledge from substantial amounts of data. Studies have shown that up to 55% of manual activities in production can be automated using AI. Deep learning (DL) is a set of algorithms to model complex and high-level abstractions in data and can better exploit large-scale datasets to build deep neural networks (DNNs). As DNNs can extract deep features from high-dimensional imaging data, DL-based approaches have achieved outstanding results in various biomedical image analysis tasks. Despite these successes in numerous other automated decision-making processes, the performance and capabilities of DNN models in complex scenarios heavily rely on rigorous feature engineering, training, and hyperparameter optimization with domain expertise. In addition, training DNN models also require substantial amounts of engineering labor and powerful hardware infrastructure (e.g., GPUs), which hinders economic and productive perspectives. Further, the unprecedented growth of biomedical devices and the variety of biomedical imaging data available today requires automated and efficient solutions. As a result, automated ML (AutoML) approaches are increasingly gaining importance and wider adoption. AutoML involves automated construction of an end-to-end ML pipeline by combining several stages of the analysis pipeline to enable domain experts to automatically build an application without requiring specialized knowledge w.r.t. statistical and ML knowledge. Although AutoML has successfully been applied to many use cases, its potentiality in cell biology has not been explored yet, mainly due to diverse datasets with multiple modalities and approaching new tasks. On the other hand, by using different kinds of metadata, e.g., properties of a learning problem, algorithmic measures, hyperparameters, or useful patterns previously learned from the data, it is possible to learn, select, or combine different learning algorithms to effectively solve a specific learning problem, often coined as "metadata'". Meta-learning, by combing AutoML techniques, has already exceeded existing learning algorithms (e.g., Google Brain's "AI building AI'" project briefly outperformed ImageNet benchmarks in 2017, according to Google). Inspired by the recent success of both AutoML and the meta-learning paradigm, we propose a meta-learning approach for automating image analysis in cell biology. In our approach, automatic learning algorithms are applied to metadata about different ML experiments in order to understand how automatic learning can be performed flexibly in solving image analysis tasks like object recognition and anomaly detection. This thesis explores if a meta-learning-based approach could help improve the performance of existing learning algorithms.

Document Actions