In May 2025, our team will participate in the IEEE International Instrumentation and Measurement Technology Conference (I²MTC 2025) in Nancy, France.
At this event, M.Sc. S.-J. Burgdorf will present recent research on the optical determination of dissolution processes in materials science.
The work addresses the automated detection of dissolution endpoints, a routine yet often manually inspected step in laboratory workflows. The goal is to reliably distinguish between fully dissolved samples and those containing residual particles, thereby improving quality control in automated laboratory processes.
To this end, a deep learning–based image classification model has been integrated into an existing laboratory automation platform, the Visual Analyzer Unit (VAU), without requiring any hardware modifications. The system employs a ResNet-50 backbone with transfer learning and classifies images into two categories: “clear” (fully dissolved) and “particles” (particles present). The decision threshold is calibrated using the Youden index on the validation set and subsequently fixed for evaluation.
Rather than introducing a novel neural network architecture, the contribution focuses on the integration, evaluation, and deployment of deep learning under realistic laboratory conditions. These include large-scale and heterogeneous datasets with varying illumination and diverse sample types. A Grad-CAM analysis reveals that misclassifications primarily occur in underexposed or low-contrast images, highlighting inherent limitations of the current imaging setup.
Compared to classical computer vision approaches relying on controlled lighting conditions or handcrafted features, the proposed method demonstrates strong robustness against heterogeneous imaging conditions and is therefore well suited for routine high-throughput laboratory automation workflows.
Come and join us at I2MTC 2026!
