Fast Protein Crystal detection and classification using Detectron2 and gaussian mixture models for continuous Bioprocessing

Abstract The fast and accurate detection of protein crystals is essential in the biopharmaceutical industry. This work presents a unified approach that integrates Detectron2-based instance segmentation with optimized image pre-processing and Gaussian Mixture Model clustering. Protein crystallization underpins structural biology and drug discovery, yet manual inspection of microscopy images is time-consuming, subjective, and error-prone. We developed an automated end-to-end pipeline for crystal detection and characterization on microfluidic platforms. The dataset contains 1,860 images, with 382 annotated, for instance segmentation. Baselines including YOLOv3 and edge-based methods, struggle in high-density conditions and produce false positives from artifacts. Models Detectron2 Mask R-CNN ResNet-101–FPN, achieving a high average statistical metric, the Dice Score of 0.90 while maintaining a false-positive rate below 5\%, as confirmed through expert validation. A GMM module in PCA-reduced features identifies four clusters with weights [0.5926, 0.1650, 0.2252, 0.0172], reflecting morphology patterns. The framework enables accurate crystal measurement and morphology analysis, achieving 92\% agreement with expert annotations and fewer than 5\% false-positive detections during validation, thereby reducing observer-dependent bias and supporting high-throughput crystallisation screening. Improves reproducibility and efficiency, supports data-driven structural biology, and rational drug design, enabling faster protein structure determination and therapeutic discovery..