Testing Precision Limits of Neural Network-Based Quality Control Metrics in High-Throughput Digital Microscopy
Chris Calderon, Dean C. Ripple, Charudharshini Srinivasan, Youlong Ma, Michael J. Carrier, Theodore Randolph, Thomas O'Connor
Digital microscopy is used to monitor particulates such as protein aggregates within biologic pharmaceutical products. The digital images that result encode a wealth of information, but much of this image data is underutilized in pharmaceutical process monitoring. For example, digital microscopy images of particles found in protein drug products typically are analyzed only to obtain particle counts and particle size distributions, even though the images also contain information that reflects particle characteristics such as shape and refractive index. However, in recent years multiple groups have demonstrated that convolutional neural networks (CNNs) can be used to extract information from images of particles comprising aggregated protein that can be analyzed to assign the likely stress at the "root-cause" of aggregation. A practical limitation of these previous CNN-based approaches is that the potential stresses inducing aggregation must be known a priori, disallowing identification of particles produced by unknown stresses. Here, we demonstrate that CNN analyses may be expanded by incorporating judiciously chosen particle standards in the recently proposed ''fingerprinting" algorithm (Biotechnol. & Bioeng. (2020) 117:3322) to allow detection of populations of particles formed by unknown root-causes from flow imaging microscopy (FIM) images. We focus on the use of ethylene tetrafluroethylene (ETFE) microparticles as standard particles that serve as a reproducible surrogate for protein aggregates. We quantify the sensitivity of the new algorithm to optical parameters such as microscope focus and refractive index changes. ETFE particles are used to explore how inherent FIM sample noise affects statistical testing procedures. When applied to real-world microscopy images of protein aggregates, the fingerprint algorithm can reproducibly detect complex, distinguishing ''textural features'' that are not easily described by standard morphological measurements used in commercially available software. Importantly, we demonstrate that the "textural features," complementing size and count information, can be statistically precise run to run (i.e., good repeatability and intermediate precision). This offers promise for both quality control applications and for detecting shifts in protein aggregate populations due to stresses resulting from unknown process upsets.
, Ripple, D.
, Srinivasan, C.
, Ma, Y.
, Carrier, M.
, Randolph, T.
and O'Connor, T.
Testing Precision Limits of Neural Network-Based Quality Control Metrics in High-Throughput Digital Microscopy, Pharmaceutical Research, [online], https://doi.org/10.1007/s11095-021-03130-9, https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=932601
(Accessed December 6, 2023)