Machine Learning-Assisted Image Analysis of Sub-visible Particles in Biologics Drug Products: A Comparative Study Using Micro-Flow Imaging and FlowCam

Sub-visible particles (SvPs) in parenteral drug products present patient safety and product quality risks. While Micro-Flow Imaging (MFI) and FlowCam are established for SvP enumeration and morphological assessments, inferring particle identity and origin from images alone remains challenging due to diverse inherent, intrinsic, and extrinsic sources, as well as stress-induced changes in particle populations. This study developed and evaluated a deep-learning workflow that integrates transfer-learning feature extraction with anomaly detection and similarity search to improve ground truth label quality and supervised classification to differentiate proteinaceous and non-proteinaceous SvPs. A comprehensive panel of SvP types relevant to biologicsincluding protein aggregates generated under thermal, pH, mixing, and agitation stresses; fatty acids; silicone oil droplets; elastomeric stopper particles; glass lamellae; and fibrous contaminantswas imaged on both FlowCam and MFI. Subject matter experts curated labels, and features extracted via VGG16/ResNet50 informed anomaly detection for label/quality control. Three classifiers (a baseline CNN, DenseNet121, and YOLOv8) were evaluated and benchmarked. Across models built with particle images of varying sizes (5−10 μm, 10−25 μm, >15 μm, and >25 μm), YOLOv8 achieved the highest performance (FlowCam: accuracy 0.973; MFI: accuracy 0.918), with strong precision/recall and specificity, outperforming both the baseline CNN and DenseNet121. Clustering analyses further revealed separations between protein vs non-protein SvPs and showed the model’s potential ability to distinguish stress-specific protein aggregates across multiple molecules, suggesting potential utility in root-cause analysis during development and manufacturing. This comparative, cross-platform study highlights the value of modern classification architectures as a complementary SvP identification tool to confirmatory spectroscopies (e.g., Raman/FTIR), thereby improving process understanding within biologics development.

Authors

Institutions

Publication Details

Journal
Molecular Pharmaceutics
Published
2026-11-19
DOI
https://doi.org/10.1021/acs.molpharmaceut.6c00682
Primary Topic
Industrial Vision Systems and Defect Detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Machine Learning-Assisted Image Analysis of Sub-visible Particles in Biologics Drug Products: A Comparative Study Using Micro-Flow Imaging and FlowCam

Chris Foti, Shikhar Mohan, Yi Li, Brooke Trimmer et al.
Molecular Pharmaceutics
Industrial Vision Systems and Defect Detection
article

Machine Learning-Assisted Image Analysis of Sub-visible Particles in Biologics Drug Products: A Comparative Study Using Micro-Flow Imaging and FlowCam

Chris Foti, Shikhar Mohan, Yi Li, Brooke Trimmer, Bowen Jiang, Shalini Unnikandam Veettil, Katelyn Esposto, Li Yi
article en

Abstract

Sub-visible particles (SvPs) in parenteral drug products present patient safety and product quality risks. While Micro-Flow Imaging (MFI) and FlowCam are established for SvP enumeration and morphological assessments, inferring particle identity and origin from images alone remains challenging due to diverse inherent, intrinsic, and extrinsic sources, as well as stress-induced changes in particle populations. This study developed and evaluated a deep-learning workflow that integrates transfer-learning feature extraction with anomaly detection and similarity search to improve ground truth label quality and supervised classification to differentiate proteinaceous and non-proteinaceous SvPs. A comprehensive panel of SvP types relevant to biologicsincluding protein aggregates generated under thermal, pH, mixing, and agitation stresses; fatty acids; silicone oil droplets; elastomeric stopper particles; glass lamellae; and fibrous contaminantswas imaged on both FlowCam and MFI. Subject matter experts curated labels, and features extracted via VGG16/ResNet50 informed anomaly detection for label/quality control. Three classifiers (a baseline CNN, DenseNet121, and YOLOv8) were evaluated and benchmarked. Across models built with particle images of varying sizes (5−10 μm, 10−25 μm, >15 μm, and >25 μm), YOLOv8 achieved the highest performance (FlowCam: accuracy 0.973; MFI: accuracy 0.918), with strong precision/recall and specificity, outperforming both the baseline CNN and DenseNet121. Clustering analyses further revealed separations between protein vs non-protein SvPs and showed the model’s potential ability to distinguish stress-specific protein aggregates across multiple molecules, suggesting potential utility in root-cause analysis during development and manufacturing. This comparative, cross-platform study highlights the value of modern classification architectures as a complementary SvP identification tool to confirmatory spectroscopies (e.g., Raman/FTIR), thereby improving process understanding within biologics development.

Molecular Pharmaceutics
Gilead Sciences (United States) (US)
Industrial Vision Systems and Defect Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon