LeukoBIAS: Analysis, mitigation, and auditing of bias in foundation model-based leukemia detection from routine diagnostic blood smears

ongoing

Foundation models have revolutionized image processing in healthcare, offering robust performance across various tasks without task-specific training. However, potential biases in these models, especially when applied to critical medical diagnostics such as leukemia detection, remain largely unexplored. LeukoBIAS aims to address this crucial gap by developing a framework for analyzing and mitigating bias in foundation model-based leukemia detection algorithms. Building on my previous work in AI-driven leukemia diagnostics, we will leverage a unique real-world dataset of over 6000 patients from my long-time industry partner, the Munich Leukemia Laboratory, to investigate biases related to sex, age, and other patient characteristics. Our approach combines advanced machine learning techniques, including multiple instance learning and attention mechanisms, with novel bias detection and mitigation strategies. The project consists of three work packages: (i) bias analysis in foundation model-based leukemia diagnostics; (ii) development of bias mitigation techniques for model fine-tuning; and (iii) exploration of intellectual property and commercialization opportunities for bias auditing. The innovative potential of the project extends beyond leukemia diagnostics. We will thus explore the scalability of our approach to other modalities and conduct a comprehensive market analysis to identify potential industry partners. LeukoBIAS will contribute to the scientific understanding of bias in medical AI and pave the way for more equitable and reliable AI-driven diagnostic tools. By addressing the requirements outlined in the EU Artificial Intelligence Act, LeukoBIAS is poised to have a significant impact on the development and deployment of trustworthy AI in healthcare. By providing a framework for bias analysis and mitigation, this project will contribute to more accurate diagnoses, improved patient outcomes, and accelerated innovation in AI-driven medicine.