An Image Analysis Study in Numbers: QDP Workflows, IT Infrastructure & Overcoming Technical Challenges

In this article, Senior Image Analysis Scientist, Chiara Asselborn shares her insight into OracleBio's quantitative digital pathology workflow, IT infrastructure, and how it helped overcome technical issues encountered in a large multiplex IF (mIF) non-small cell lung cancer (NSCLC) study.

—

When someone asks me as an Image Analysis (IA) Scientist which studies I like best I would have to say multiplex immunofluorescence (mIF) studies. Yes, they are also the most challenging studies, but they leave room for creative problem-solving, which I thoroughly enjoy as an IA Scientist.?

However, the only way it is possible to overcome the challenges often encountered within complex mIF studies is through the powerful image analysis workflows that we have in place at OracleBio.?

It all starts with our excellent IT infrastructure. At OracleBio, we migrated our image analysis workflow from?mostly physical (on-premise) hardware onto our new purpose-built, high-performance cloud environment built on Amazon Web Services (AWS). This included our entire image repository as well as the image analysis software we use day-to-day; Indica Labs HALO and Visiopharm.?

Not only did this help with more efficient algorithm loading and internal processing times, but it also significantly improved the viewing experience of images and results for our Clients in HALO Link, and our own proprietary viewing platform, ‘Observer’. What’s more, we gained the ability to more closely interact with Client pathologists and IA teams around the workflow, for example in creating tissue annotations or reviewing algorithms.

OracleBio’s IT Infrastructure

In this example of a 7-plex mIF non-small cell lung carcinoma (NSCLC) study in collaboration with Bristol Myers Squibb (BMS), we successfully analysed n=96 mIF images, with an impressive accumulated image storage size of 588 GB.??

To be able to cope with this large amount of data, our IT infrastructure in AWS is set up with scalability in mind. This means that we can launch and shut down up to n=400 CPU servers and n=75 GPU servers at any one time?if required.

This introduces great flexibility and has significantly improved image processing times, especially for large studies such as this one. This has allowed us to save many hours of project time, and importantly, get the data back to our client in record time.

Taking advantage of this setup, it was thus possible to efficiently process a sequence of n=3 Deep Learning (DL) based detection apps (n=2 tissue classifiers + n=1 cell detection app), n=1 non-DL based area quantification app and n=11 separate spatial analysis runs for each of the 96 images.

The apps were then run in 3 regions of interest: tumor, stroma, and invasive margin.?That’s a total of n=1440 analysis runs performed for this study.?

To put this analysis into context further,?we identified an average of n=90,582 cells per image. Across the n=96 images analyzed, we identified and returned data for an impressive n=8,514,676 individual cells.??

Client, Pathologist, and CRO interaction

Great IT infrastructure is crucial, but it is only part of the story.?

For me as an Image Analysis Scientist, one of the most exciting developments at OracleBio in the last couple of years was the implementation of Deep Learning (DL) based analysis algorithms into our analysis workflow. We have Artificial Intelligence capabilities for both our software, Indica Labs HALO and Visiopharm, and use DL as standard for most of our algorithms. This gives us the flexibility to execute even the most complex analysis requests and deal with staining variability between images, which is common for mIF studies.

The n=3 apps used in this study were therefore all developed using DL. A typical example of a DL app is the tissue classifier developed for this study to segment tumor and stroma while excluding necrosis and any obvious artefact.?

Our workflow also allows us to investigate more specialized algorithm requests and can provide the client with valuable extra granularity for their analysis readouts and interpretation of the data. For this study, following an in-depth discussion of the Client’s requirements, we agreed to develop a classifier that identified and excluded macrophages specifically residing in alveolar spaces of the lung tissue. Advice and input from our internal clinical pathologists are essential in the development of these classifiers, which further complements our continuously expanding image analysis expertise.

Expertise in Image Analysis and Pathology?

This brings me back to my initial point about enjoying the creative problem-solving aspect of complex mIF studies.?

All studies are different and while some challenges are commonly observed in many mIF studies (see my colleague Cristina Suanno’s recent blog here about common issues with autofluorescence), other challenges can be unique to a study or client.?

By deploying an image analysis workflow that combines the following; a state-of-the-art IT infrastructure, image analysis expertise, AI capabilities within 2 different software, and an ability to collaborate extensively via HALO Link and OBserver, we can start to answer complex morphological and spatial-related questions for our clients. This is what I find most exciting about working at OracleBio.??

If you would like to know more about how we addressed this NSCLC study in particular, please check out the previously published poster on our website and keep an eye out for the second part of this blog, which will go into more detail about the science behind this study.?

If you would like to discuss how OracleBio can support your image analysis studies, you can arrange a chat with our team here.

This post is written by OracleBio Senior Image Analysis Scientist, Chiara Asselborn.