Why isn’t subject matter expert (SME) analysis of images or basic software assisted SbVP image classification an effective solution?
When we meet with customers to discuss our technology, we are often asked if we can parse particle images based on a qualitative set of human interpretable features – like silicone oil, air bubbles, etc. Even an SME can only visually resolve gross features in the amorphous, often blurry images. We could easily deliver a solution that classifies these images by human defined type – but it would likely only tell you what you already know: that your sample contains air bubbles, silicone oil droplets, SbVPs that are likely protein aggregates, silicone oil + protein particles and aggregates. These image classifiers are only as good as the data fed to teach the classifiers. When they encounter an unknown particle, it is forced into its nearest classification category. (Figure 1)
SbVPs all have morphological and textural features that tell stories about their origins, different drug formulations and stress conditions result in different features. Those tiny feature differences are virtually impossible for humans to view and characterize. Even if SME is very good they can’t reliably see below 50 µm, let alone discern the embedded subtleties contained in the image. 50 µm is an order of magnitude larger than particle sizes of concern in a biologic drug product today. If your SMEs create defect libraries based on their qualitative analysis of these images, use of the library is limited by human perception of those features. Add fatigue and variations in SME perceptions, and grouped populations are even more qualitative, virtually ensuring validation of the method is unachievable. Further, these libraries are limited to the SbVPs that your SMEs have seen and classified. If you employ an ML classification algorithm to these libraries for analysis of particles from other samples, the particle you’ve never seen or classified is grouped into the nearest class match – which may not be a match. These simple classification solutions, whether human or software assisted, even when based upon quantitative data like aspect ratio, circularity, or longest dimension, must put unknown SbVPs into a class – even when it isn’t a match. These ML classifiers are task focused on forcing a particle image into a known, labeled class, not designed to watch for the unique, new feature or feature variations, or new outlier particle classes.
Our approach looks at the heterogeneous mix of particles (aggregates, SbVPs, air/gas bubbles, silicone oil) in a sample. By employing well-designed experiments and looking at the entire sample, we simplify sample data preparation, and deliver statistical tools to identify shifts in any of these constituent populations, or identify outlier particle populations not previously seen. There is definitely a time in drug development to understand exactly what each particle is, its origin and risk – and those tests will be discussed at length in the next article in this series.
Figure 1
The parsing question stems from wanting to see and understand every SbVP in every sample, rather than using statistical methods to detect a shift in the normal particle population. We call this method “seeking the needle in the haystack”. This approach doesn’t challenge that old identify-every-particle-paradigm or leverage the rich textural and morphological features contained in the images that are not detectable by human SMEs. DML allows us to leverage these features and still identify statistical variations in particle features – the new features or characteristics associated with outlier particles. Our process quantifies all the particles in your sample – the entire heterogeneous mix of silicone oil bubbles, air/gas bubbles, protein aggregates, SbVP’s, etc. Instead of classifying all the needles in the particle haystack, we characterize the haystack. Our approach allows us to deliver quantitative data and demonstrate when a statistical shift has occurred in your haystack or particle population. (Figure 2)
Figure 2