Biomedical image analysis, especially in cell and tissue microscopy, is a necessary and tedious task in bio laboratories that requires biomedical expertise and is therefore highly cost-intensive. Additionally, the subjectivity of the analysis and the susceptibility to human errors affect the comparability of scientific results. While deep learning holds promise for automating these analysis tasks, biomedical specialists rarely possess the necessary skillset and capacities to develop, deploy and maintain robust DL applications for their individual analyses. We propose a four-stage domain-specific image automated machine learning (AutoML) system architecture that aims to balance the generality-specificity trade-off that AutoML systems typically suffer from and apply it to the domain of cellular image analysis. We introduce AIxCell to automate the design, construction and training of deep learning-based pipelines for cellular image analysis. By leveraging a portfolio-based meta-learning approach and multi-fidelity training (i.e., successive halving), AIxCell identifies, trains and provides the optimal image analysis pipeline to the biomedical expert. The results show the effectiveness of meta-learning in a domain-specific setting and that AIxCell reliably outperforms baseline solutions. Our findings highlight the potential of domain-specific image AutoML systems to enhance efficiency and limit cost in biomedical image analysis.
Roberg et al. (Mon,) studied this question.