Project Published on

Project - Automated Biomarker Quantification for Microscopy Images

Authors
  • Name
    Twitter

Project Overview

TWA's Biomarker Quantification Project is in collaboration with Spotlight Pathology.

Spotlight provided the data and Microscopy/biological expertise and TWA provided computer vision capability and rapid development to create this addition to Spotlight's production pipeline.

The project involved TWA learning some immunohistochemistry (very new to us!) and close consultation with Spotlight's domain experts to achieve a successful prototype.

Biomarker Quantification from Microscopy Images

Note that a typical microscopy slide is only 3 inches by 1 inch. However whole slide image scanners are used to extract an extremely high resolution image of the whole slide, that can have resolutions close to a micron.
The resulting data can be very high dimensional. For example 100,000 pixels by 250,000 pixels.

Example whole slide image An example section of a whole slide image.

Viewing the whole slide at a cellular and structural level is like zooming in and out in Google Maps, but the map is only a very small part of human anatomy.

Manual biomarker quantification either requires counting cells manually or making an eye-ball estimate of the percentage of biomarker-positive cells in a sample. As such, manual assessment is time consuming and can suffer from high levels of subjectivity.

Automatic Diagnosis

The main aim of the project was to develop an automated AI tool capable of evaluating each microscopy slide to quantify biomarkers to inform cancer diagnosis.

Spotlight had a set of objectives for the AI tool as follows:-

  • Efficiently identify informative regions of the slide
  • Find all visible cells in each region, including nuclei and cytoplasm
  • Compute relevant cell statistics for each region. For example, the proportion of cells which are biomarker positive.
  • Accumulate all the evidence into a report that the pathologist can review.

TWA in collaboration with the Spotlight were able to quickly develop a successful prototype which had very strong correlation with the manual (ground truth) diagnosis, as determined by expert pathologists. The AI tool can produce extensive quantitative results over the whole test region (instead of isolated patches) and was able to compute a diagnosis much quicker than a human simply viewing the slides.

Note that the intention was to develop a tool that aids the pathologist and saves them time on tedious manual tasks such as cell counting. The final result is still accessed by a pathologist who can incorporate the automatically computed quantitative biomarker results with their own qualitative findings and other relevant patient information into a final written report.

Summary

If you have a computer vision/medical AI project you wish to develop, contact us for a free consultation/discussion on how TWA can help you in your use case. TWA have extensive experience in computer vision and software as a medical device products, that are developed and deployed under ISO13485 guidelines.