This work presents an open-source workflow for automated malaria parasite detection from low-cost microscopy images, addressing both a technical challenge in biomedical AI and the broader theme of digital sovereignty. In many biomedical contexts, reliable AI tools depend on expensive acquisition devices, proprietary software, or infrastructures that are difficult to reproduce and adapt locally. Low-cost microscopy can improve accessibility, but it introduces a domain shift: models trained on images acquired with high-cost microscopes often show a measurable performance drop, for example in terms of mAP50, when applied to images acquired under less controlled conditions.
We use YOLOv3 as the reference model for automatic parasite detection. As a baseline, the model is trained on high-cost microscope images and evaluated on low-cost microscope images, in order to quantify the impact of domain shift. We then investigate preprocessing techniques and targeted data augmentation strategies to reduce this performance gap without modifying the model architecture.
The preprocessing pipeline includes denoising, luminance and color correction, local contrast enhancement, and sharpening operations. Filter selection is performed through a greedy strategy, while filter hyperparameters are optimized using Optuna. In parallel, the training process is enriched with targeted data augmentation, applying transformations that reflect those later used on low-cost microscope images at inference time.
The complete workflow is implemented in KNIME, an open-source, low-code analytics platform that enables reproducible data processing and machine learning pipelines through a visual interface. The workflow is designed as part of a community-oriented KNIME extension, allowing researchers and practitioners to inspect, adapt, and reuse the pipeline without relying on proprietary software components.
Experimental results show that preprocessing alone does not improve performance, whereas its integration with targeted data augmentation yields an improvement over the baseline. Although the gain is modest, the results demonstrate how lightweight, transparent, and open techniques can support more accessible biomedical AI, contributing to reproducibility, local control, and digital sovereignty in healthcare-oriented applications.
