AIxCell – Democratizing Deep Learning for Biomedical Image Analysis

Overview

As part of my role in the AIxCell research project—funded by the German Federal Ministry for Economic Affairs and Climate Action—I contributed to building an AutoML platform that empowers biomedical professionals to independently design, train, and deploy deep learning models. The platform addresses a critical pain point in biomedicine: the lack of accessible, scalable tools for image-based analysis of complex biological structures.

By automating traditionally manual, expertise-heavy tasks such as cell segmentation and classification, AIxCell drastically reduces time, cost, and reliance on external AI experts—while maintaining scientific rigor and reproducibility.

Problem Solved

Biomedical image analysis is typically manual, error-prone, and domain-specific. Custom AI solutions are expensive and inaccessible to most laboratories, while internal teams often lack the machine learning expertise to build reliable models. As a result, many biomedical research workflows suffer from low reproducibility, poor scalability, and inconsistent results across devices and experiments.

Solution & Strategic Impact

Lowered the barrier to entry for deep learning in life sciences with a self-service AutoML tool.
Enabled biologists, virologists, and clinicians to independently train and use deep learning models—without writing code.
Reduced dependency on external consultants or specialized AI teams, cutting both time and cost.
Standardized and scaled image-based workflows for more reproducible research outcomes.

Key Innovations

Meta-Learning Engine (AutoKonfig): Automatically selects optimal algorithms based on task, data, and available resources—continuously improving via feedback loops.
Modular Deep Learning Library: Includes pre-trained models with rich metadata for rapid task-specific adaptation.
User-Centric Design: A fully integrated GUI allows domain experts to upload data, annotate images, define tasks, view results, and deploy models—all in one place.

Business & Research Impact

De-risked AI adoption in biomedical institutions lacking in-house ML teams.
Reduced cost of AI model development by minimizing need for custom outsourcing.
Accelerated research cycles, enabling faster iteration and validation of hypotheses.
Increased reproducibility and comparability across labs and experiments through standardized, automated pipelines.

My Contribution – Master's Thesis & Research Publication

As part of this project, I authored my master’s thesis titled:
“Benchmarking of Deep Learning Algorithms for Stem Cell Classification”

Highlights:

Developed a three-stage image analysis pipeline (preprocessing, modeling, postprocessing) for cardiomyocyte segmentation in confocal microscopy.
Conducted 173 experiments benchmarking 18 encoder-decoder architectures for small dataset performance.
Achieved test accuracies up to 82%, demonstrating the feasibility of deep learning in low-data, high-complexity biomedical tasks.
Contributed findings to the MIDL 2022 conference in Zürich, further validating the real-world value of our automated pipeline.

This work directly fed into AIxCell’s model library and meta-learning logic, enhancing the system’s performance in cardiomyocyte analysis use cases.

Technical Highlights

Domain-specific AutoML for semantic segmentation and classification
Modular pipeline design for reusability and flexibility
Meta-learning approach for continuous optimization of model selection
Focus on low-data performance and real-world constraints of lab environments

Tech Stack

Deep Learning Frameworks: TensorFlow, PyTorch
Image Processing: OpenCV, Scikit-image
AutoML & Meta-Learning: Custom-developed Python modules
Deployment: Containerized apps via Docker, GUI application

Funding & Collaboration

The AIxCell project was funded through the AiF / IGF program under grant 21361 N by the German Federal Ministry for Economic Affairs and Climate Action. It was developed in collaboration with Fraunhofer IPT and various academic and industry partners.