TOXBOX project partner Steinbeis Advanced Risk Technologies Institute doo (SARTIK) is developing automated deep learning systems for cellular response classification in toxicity screening, in collaboration with Empa. Continue reading to get a deeper insight into their work.
Research Objectives
Within the TOXBOX project, we are developing automated classification systems for microscopic images to assess cellular toxicity responses. Our work addresses a critical bottleneck in toxicological screening: the labor-intensive manual analysis of cellular changes following exposure to potentially harmful substances.
The system classifies high-resolution microscopy images into three cellular states: control (healthy), irritation, and sensitization responses. This automated approach provides consistent, rapid assessment capabilities to support regulatory toxicology and safety evaluation processes.
Methodological Approach
Working with limited training data – a fundamental constraint in specialized toxicological datasets – we developed innovative solutions to optimize model performance. Our dataset comprised images from 14 experimental conditions (Kera48-Kera64) with three classes – control, irritation and sensitization images.
To address the inherent data imbalance (352 control vs. 123 irritation vs. 121 sensitization images), we implemented intelligent strategies that generate multiple informative regions from each source image. This approach created a balanced training dataset, effectively transforming data scarcity into a methodological strength.
Technical Implementation
The classification system employs DenseNet121 architecture with transfer learning, fine-tuning pre-trained ImageNet weights for toxicological classification tasks. This approach leverages existing learned features while adapting to domain-specific cellular morphology patterns.
Soft attention mechanisms provide interpretability through visualization of decision-critical image regions. These attention maps confirm that the model focuses on biologically relevant cellular structures, validating the system’s learning of fundamental toxicity markers rather than spurious correlations.
Current Challenges and Future Directions
Our immediate research focus involves expanding the classification system into an integrated web-based platform for comprehensive toxicity analysis. The platform will enable researchers to upload microscopy images and receive immediate AI-powered assessments of cellular responses.
Key technical challenges include improving sensitization classification accuracy and incorporating multimodal data integration. The ultimate objective is developing a robust AI system that augments expert toxicological assessment, providing accessible, consistent analysis while enabling specialists to focus on complex cases requiring human expertise.
Impact and Vision
This work represents a paradigm shift toward democratized access to expert-level toxicological analysis. By combining machine learning expertise with toxicological domain knowledge, the TOXBOX project advances automated safety assessment capabilities, making toxicity evaluation more accessible and comprehensive for the broader research community.