A tube that looks clear contains particles. Between roughly 1 and 100 micrometres sits a population large enough to be counted one at a time yet too small to resolve by eye, and this is the band where two routine instruments — light obscuration and flow imaging microscopy — are asked to produce a number. Run both on the same tube and they can disagree by a factor of ten or more, and the disagreement is systematic rather than random. Knowing which direction it runs, and why, is the difference between a subvisible count that constrains something and a count that reflects which instrument happened to be free.
What each instrument actually records
Light obscuration is a single-particle optical sensor. Sample is drawn through an illuminated cell; each passing particle attenuates the beam reaching a photodiode, and the amplitude of that signal drop is converted to an equivalent diameter through a calibration curve, conventionally built from polystyrene spheres suspended in water. The pharmacopoeial subvisible test that the technique underwrites requires counts at the two thresholds of 10 micrometres and 25 micrometres [1]. What comes out is a count in a size bin. It carries no information about shape, composition or origin.
Flow imaging microscopy pushes sample through a flow cell in front of a digital camera and stores an image of every particle in the illuminated volume. In the method development reported for glatiramer acetate, a polypeptide, flow imaging quantified particles across a 1–500 micrometre range, and accuracy, precision, linearity and specificity were assessed for that method [2]. Because the primary record is an image rather than a pulse height, the output is a per-particle table of size and morphology. Micro-flow imaging was described as sensitive to transparent protein particles and as permitting several particle populations to be analysed independently in a single run even when silicone droplets and air bubbles are present alongside proteinaceous material [3].
That difference in what is recorded — a scalar versus an image — is the root of everything that follows.
Which way the disagreement runs
With standard instruments the direction is reproducible. Compared head to head, the two techniques gave similar results on polystyrene standards; in a protein-containing solution, an immunoconjugate, micro-flow imaging indicated significantly more particles than light obscuration. The authors attributed the difference to imaging technology detecting non-spherical particles and particles with refractive indices similar to the solvent [1]. Across several therapeutic protein injections, flow imaging counts were again much higher than light obscuration counts, and the size of the gap depended on the product and on the features of the particles present; some preparations contained large numbers of transparent, elongated particles that could escape detection by light obscuration [4].
The mechanism is contrast. Obscuration measures how much light a particle removes from the beam, which depends on the refractive index difference between particle and medium. A determination of the refractive index of protein particles from two different proteins returned a value of 1.41 — close enough to concentrated aqueous formulations that the contrast can collapse. Raising the refractive index of the formulation, through high protein concentration and/or sugars at pharmaceutically relevant levels, significantly underestimated the subvisible particle concentration determined by light obscuration and by micro-flow imaging, and a refractive index match caused particles to become undetectable by either technique [5].
The refractive index story is the one most often told too simply. The convenient shorthand — obscuration is refractive-index dependent, imaging is not — is wrong in its absolute form. Flow imaging is less affected, not unaffected. The same study recommended polytetrafluoroethylene particles as a probe for testing whether a given formulation produces refractive index effects at all, and light-independent techniques such as resonant mass measurement where such effects are present [5].
The magnitude has been quantified directly. Before any correction, results from the three instruments used — two flow imaging and one light obscuration — disagreed by a factor ranging from 3.1 to 48 in particle concentration over the size range from 2 to 20 micrometres. Applying bias corrections — refractive index characterisation and scattering models for light obscuration, silica bead calibration in water–glycerol solutions for flow imaging — brought an average factor of 14 down to an average factor of 1.5 [6]. Two conclusions follow. The raw disagreement is mostly a known optical bias rather than an unknowable difference in what the instruments see, and correcting it requires information about the particles' optical properties that a routine count does not have.
Calibration is not a neutral step
Because the size axis is defined by a calibration curve, the choice of calibrant is part of the measurement. Substituting lower refractive index silica beads suspended in sucrose solution for the conventional polystyrene-in-water suspension showed that sizing accuracy was strongly dependent on the properties of the reference material used for calibration; flow imaging microscopy suffered from the same artefact, to a smaller extent. Across different light obscuration sensors and instruments, sizing accuracy also varied from instrument to instrument, strongly depending on the properties of the sensor. The authors drew the practical conclusion that light obscuration and flow imaging microscopy are not truly orthogonal analytical methods, and noted that matching the optical properties of calibrant and test item improves sensitivity but has poor practical applicability because real samples contain multiple particle types at once [7].
Nor is the deficit intrinsic to obscuration as a physical principle. Standard light obscuration has limitations for translucent proteinaceous particles and for particles at and below about 2 micrometres, but a modified sensor that monitors the length (width) of the signal rather than its height (amplitude) improved detection below 2 micrometres, reduced artefacts when measuring low concentrations of translucent protein particles, and gave higher counting accuracy than both flow imaging microscopy and standard light obscuration in that evaluation [8]. That is a modified sensor, not a routine one, and it should not be read as a general ranking of the two techniques — but it does mean the blindness of light obscuration to translucent material is a property of the conventional readout, not of the physics.
How reproducible is any of this
An interlaboratory comparison run across 23 laboratories from industry, government and academia circulated a polydisperse suspension of abraded ethylene tetrafluoroethylene particles, chosen because they closely mimic the optical contrast and morphology of protein particles, and asked for sizing and counting from 1 to 25 micrometres. Where the data sets were restricted — lower diameter limits from 1 to 5 micrometres for light obscuration, a specified manufacturer and instrument setting for flow imaging — relative standard deviations were approximately 25% and approximately 30% respectively. Unrestricted, the reported counts carried relative standard deviations of approximately 50% for both. In the reported counts, differences between instrument manufacturers were not statistically significant for light obscuration but were significant for flow imaging [9] — a different endpoint from the sizing accuracy discussed above, and not in conflict with it.
Read that as a floor on what a between-laboratory comparison can resolve. A two-fold difference in counts reported by two laboratories on nominally identical material sits inside the spread of an unrestricted protocol. If a comparison matters, fix the instrument, the settings and the size cut-offs before the samples are drawn, not afterwards.
The band below two micrometres
Neither routine method covers the submicron region well, and the morphological discrimination that makes flow imaging useful degrades before the counting does. Flow imaging microscopy is successfully applied to differentiating protein aggregates from silicone oil droplets above about 5 micrometres on the basis of morphological appearance. Combining oil-immersion flow imaging microscopy with convolutional neural networks extended that discrimination to between 0.3 and 5 micrometres, differentiating heat-stressed therapeutic monoclonal antibody aggregates from artificially generated silicone oil droplets with misclassification rates of about 10%; particles in the low submicron range were particularly difficult, because their morphological appearance becomes very similar [10].
Below that, the field falls back on nanoparticle tracking analysis and resonant mass measurement, and their agreement is poor. A cross-industry survey of 52 unique protein products covering 62 dosage forms found measured particle concentrations spanning four orders of magnitude by nanoparticle tracking analysis and three orders of magnitude by resonant mass measurement, with the concentrations determined by the two techniques differing significantly for both control material and actual product. Their results suggested these techniques exhibit higher variability than well-established subvisible particle characterisation techniques such as flow imaging or light obscuration, and the authors held that in their current states the techniques can be used for product and process characterisation but may not be suitable for release or quality control testing [11]. This connects directly to what dynamic light scattering can and cannot tell you: an intensity-weighted hydrodynamic size distribution is a different observable again, and none of these three counts particles the way the 10-micrometre bin is counted.
Handling manufactures the measurand
Every operation between the container and the sensor can create particles. Vacuum degassing of a reconstituted freeze-dried etanercept solution — a fusion protein — caused a substantial increase in subvisible particles, detected by micro-flow imaging and by dynamic light scattering. The proposed explanation was that bubble collapse under vacuum produces substantial heat and ultimately free radicals, but adding the free-radical scavenger ascorbic acid had minimal effect on particle formation, so radical formation was probably not the main cause [12]. Degassing is a routine preparation step included precisely to prevent air bubbles being counted as particles, and here it generated the analyte.
Withdrawal does the same. Repackaging human immunoglobulin and lipid nanoparticles from a vial into plastic syringes generated microparticles, with greater numbers using silicone-oil-lubricated syringes than silicone-oil-free syringes, and continuous accumulation across repeated repackaging cycles for both formulations. Filter-integrated needles, stabilisers such as acetyl-arginine and Tween 20, and alternative withdrawal techniques each reduced the effect [13].
Two protocol consequences. First, run the diluent and buffer through the same pipette tips, syringes and tubing you use for the sample, and treat the count you report as the difference; an unblanked count is an instrument reading, not a property of the material. Second, remember that dilution changes two things simultaneously. It lowers particle concentration, and it changes the refractive index of the suspending medium, which changes the contrast on which both instruments depend [5]. A dilution series that does not scale linearly is telling you something about optics as much as about coincidence.
Assigning particles to a population
A count is not an identification. Three populations coexist in the same size band: intrinsic proteinaceous particles arising from the material itself, extrinsic contamination introduced from outside, and silicone oil droplets from lubricated surfaces — alongside air bubbles, which are a handling artefact rather than anything in the sample [3]. Silicone oil droplets are near-perfect spheres; proteinaceous particles are typically irregular and translucent, and that asymmetry is what image-based classification exploits. A combination of four morphological image parameters was found to provide a reliable basis for automatic distinction between silicone oil droplets and protein aggregates, with low misclassification errors [14].
Robustness across sample types is the harder problem. Automated approaches for silicone oil discrimination were still described as lacking robustness in accuracy and transferability when a two-step image-based filter was developed that is trained on silicone oil droplet images only, and therefore independent of the protein sample being imaged. Benchmarked against alternatives it gave the best overall performance, with accuracy particularly good for higher-resolution images, and it distinguished silicone oil droplets in protein solutions that were not used to create the filter [15]. The design principle generalises: a classifier trained on one protein-plus-oil pairing carries that pairing's assumptions with it.
What image classification licenses, then, is a class assignment with a stated error rate — roughly 10% between 0.3 and 5 micrometres in the oil-immersion work [10] — not a chemical identification. Light obscuration licenses no class assignment at all, which is why it also mis-reports the silicone population: light obscuration underestimated both the size and the number of silicone oil droplets in prefilled syringe products relative to flow imaging [4].
Why the SEC trace is a different measurement
Size-exclusion chromatography reports soluble oligomeric species; subvisible counting reports insoluble particles. They are not two views of one quantity. In the degassing work, size-exclusion HPLC could not detect any change while micro-flow imaging and dynamic light scattering both registered the increase in subvisible particles [12]. A flat SEC trace does not license the conclusion that nothing aggregated, and a stable subvisible count does not license the conclusion that no soluble high-molecular-weight species formed.
The link between the two assays is also perturbed by the step that usually precedes both. Loss of high-molecular-weight species was observed during sterile filtration of eight monoclonal antibodies formulated at low protein concentration across a hydrophilic polyvinylidene fluoride membrane; the adsorption was specific to the high-molecular-weight species, differed in degree between antibodies, and was concluded to arise from electrostatic interactions with the filter surface, being reducible by increasing the ionic strength of the buffer [16]. A sample that has been filtered has had part of the population it is about to be assayed for selectively removed. That belongs in the method description, not in the discussion.
The defensible position is narrow and worth stating plainly. A subvisible count is a count of particles above a stated diameter, measured by a named instrument, on a named calibration, in a specified matrix, after a specified handling sequence. Change any one of those and the number changes. Report all five alongside the count, and it constrains something; report the count alone, and it does not.