Quality Bioimage SEGmentation - Extending Allen Cell and Structure Segmenter software and disseminating AI templates and testing suites for quality assurance and reusability in a broader bioimaging community

• Applicant: Dr. Jianxu Chen
• Subject Area: Cell Biology ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 528777169

The Allen Cell and Structure Segmenter (ACSS) is an open-source Python software for intracellular structure segmentation in 3D fluorescent microscopy images. It seamlessly combines (1) classic image segmentation workflows with a large collection of recipes for organelles of various morphologies and (2) iterative deep learning (DL) workflows without needing tedious manual annotations. The software was originally built for 3D integrated analysis of 25 different cell organelles permitting the discovery of the fundamental principles by which stem cells organize themselves. Shortly after the first software release, it has been adopted by the cell biology community for various problems and applications. In general, ACSS was developed following basic software engineering standards, but there is still large potential for improvement. In this project, we aim to improve its quality and maximize its impact in the broad microscopy image analysis and cell biology community from three aspects. 1. Quality Assurance: We find that most of common open-source DL-based bioimage analysis software (in PyTorch) has marginal or no proper testing, even the basic code testing. So, after first establishing clear quality metrics and necessary automation tools to monitor the software quality, we will focus on developing systematic code/data/model testing for the DL part of ACSS, which we aim to spin out later as a handy testing suite for DL-based bioimage analysis systems. 2. Code Reusability for further development: Nowadays, you could find many different implementations of a simple U-Net model in different bioimage analysis software, even in the same lab. We plan to refactor the DL components in ACSS with the new AI template concept from our recent work in order to improve code understandability, reusability and extendibility. Such template-like modularized DL components will not only make systematic testing more viable, but also stimulate more code re-use and further development. 3. Usability and impact: According to our community users, ACSS is uniquely effective but could be more useful if we could extend the scope of applicability and provide better graphical interface for iterative DL workflows. So, we plan to further genericize the original functionalities to enable wider applicability (e.g., supporting 2D), and collaborate with the Allen Institute for Cell Science to wrap the iterative DL workflow as a napari plugin to make it more accessible. Through cooperation, e.g. as a partner of the NFDI4BioImage consortium, we will bring this quality-assured, user-friendly bioimage analysis software to the community, to developers and users, for bigger impact. In this project, the planned quality improvement for ACSS will not only improve the quality of this software itself, but also serve as an important step (e.g., comprehensive testing suite of DL and AI templates for bioimage analysis) towards building a community with better DL-based bioimage analysis software.

• DFG Programme: Research Grants
• International Connection: USA
• Cooperation Partner: Dr. Susanne Rafelski