Impedance-based particle tracking at micrometer scale with limited training data: a deep learning architecture...

In this study, we present and analyze a deep learning framework designed to achieve real-time particle localization from successive electrical impedance measurements with a limited training dataset. The goal is to provide the research community with a comprehensive analysis of the role of the different constituent elements of a network and a way to optimize their parameters. All the code is made available (see SI). The proposed algorithm combines a temporal convolutional network (TCN), a bidirectional long short-term memory (BiLSTM), a fully connected network (FCN), and an automated parameter tuning engine. Although vision-based systems could be used to track the position of micrometer-sized particles, this modality has practical limitations due to the size of the equipment, its complexity, and the narrow field of view. Thus, an emerging approach for localization is to exploit the correlation between electrical impedance measurements and the position of the successive particles. While physics-based algorithms have been proposed for localizing particles from impedance measurements, they rely on analytical assumptions on feature correlation. Learning-based approaches remove this bias by directly mapping patterns from raw data. However, state-of-the-art models typically require massive training datasets and offer minimal transparency regarding parameter selection. To contribute to the deployment of such algorithms in the lab-on-a-chip community, we introduce a framework that unifies feature extraction with a data-driven architectural tuning engine, designed to maintain predictive capacity under data-scarcity constraints. This paper analyzes the role of each element of the algorithm with respect to the particularities of particle localization, with the data made available alongside the code. The methodological robustness of the framework is validated via a micrometer-scale particle localization case study. Using a small training dataset of only 53 polystyrene beads of 8 μm and 5 μm, the optimized model establishes a localization accuracy of 3.34 μm ± 1.24 μm, while demonstrating real-time prediction times as low as 0.58 ms to support its potential for closed-loop control applications. While this work validates the methodology using uniform beads to evaluate network behavior independently of biological variability, the underlying optimization framework is inherently generic and so, applicable to more complex, heterogeneous biological samples, and will contribute to the deployment of data-driven approaches in the community.