The DaedalusData workflow supports product quality engineers in exploring large collections of particle contamination images from in-vitro diagnostics (IVD) consumable manufacturing, labeling particles by material type and visual characteristics, and externalizing tacit domain knowledge through label-informed projections for data-driven quality decision-making.
The process begins with input and preprocessing, where a collection of particle images (thousands of microscopy photographs of contaminant particles found on IVD consumables), each associated with computationally extracted image-context attributes (six size measures: area, perimeter, Feret diameter, equivalent diameter, convex area, filled area; and three shape measures: circularity, elongation, convexity) and production-context attributes (supplier, lot number, production date), are normalised and one-hot encoded to produce a unified feature vector per particle suitable for dimensionality reduction.
An initial exploration view is generated through two arrangement modes. In the Attribute View mode, particles are arranged in columns partitioned by a chosen single attribute with vertical ordering by elongation. In the Projection View mode, the expert selects multiple attributes and a UMAP dimensionality reduction is computed, positioning similar particles close together. Either mode produces a canvas visualisation displaying particle image thumbnails at their computed positions, with interactive zoom, pan, and visual customisation to manage overplotting.
The expert enters an iterative exploration and labeling loop, inspecting the canvas to identify areas of interest—clusters of visually similar particles, anomalous groups (e.g., unusual colour concentrations in a lot), or patterns relating particle appearance to production context. They apply attribute-based filters to narrow the search space and use lasso or rectangle selection to isolate a particle group. A selection-inspection view provides summary statistics for the selected particles, enabling the expert to assess whether the selection represents a coherent, meaningful group. If so, the expert defines a label (e.g., "blue", "bright", "fiber") and assigns it to the selection. If not, they refine the selection or adjust filters. This cycle repeats until the expert has identified and labeled the principal particle types of interest.
Once a substantial number of particles have been labeled, the expert proceeds to knowledge externalization through label-informed projections. Labels are organised into structured label alphabets representing domain perspectives (e.g., a "Material Type" alphabet). These label alphabets are then used as augmented attributes: the expert selects a combination of original attributes and label-based attributes, and a semi-supervised UMAP projection is computed that incorporates the label semantics into the particle positioning. The resulting label-informed projection reveals new relationships—how labeled groups relate to production context, which unlabeled particles cluster near existing labels, and whether the current labeling captures the data structure well. The expert identifies newly revealed patterns, labels additional particles, and recomputes the projection with updated labels. This cycle repeats until the expert judges that sufficient knowledge has been externalized.
Finally, in a knowledge synthesis and decision-making step, the expert compiles findings from the exploration and externalization process: confirmed contamination patterns (e.g., blue lint particles traced to changed work clothing), supplier quality assessments (delivery homogeneity across lots), lot disposition decisions (acceptable vs. to be discarded), and recommendations for process improvements. The augmented dataset with expert labels is persisted for collaborative use and downstream data-driven decision-making.