Today, depending on the setting, smear microscopy will have a sensitivity range from 40%-60%, leaving far too many TB patient cases undetected. Undetected patients return to the general population and ensure the ongoing transmission of the disease. It is not that the smear microscopy technology is deficient, but rather it is constrained by the human-vision side of the analysis. Human vision has limitations. It gets tired, distracted, and unable to perform at consistently high levels over long periods of time. Inspecting stained smears using a microscope is further limited by the process of adverse selection. When 90% of all slides inspected are negative and a slide presents with a minimal number of TB rods, it is likely to be missed. The detection challenge increases as the workday extends or the technician is new to the position. The inherent challenges in smear microscopy create the ideal environment for our computer-vision, using the best of image processing science and digital technology to increase sensitivity and diagnostic consistency.
Using Computer-Vision to Detect a One Atom Difference
Machine learning technology used in training our software to detect a military-grade explosive is the same machine learning technology used to train software to detect tuberculosis, malaria, or breast cancer. To more clearly illustrate the strengths of our computer-vision capability used in TBDx™, and how it can supplement human-vision, we often use an example from our explosive detection experiences.
In the pictures below you see on the left two X-ray images. To the human eye they are indistinguishable from each other. To the right of each image is our computer-vision analysis of these same images. Notice the differences? Though human vision has no effective way to interpret these X-ray images, our computer algorithms can analyze these pixels, extract features, and classify the top left image as water (H2O) and the lower left image as a 35% concentrate of hydrogen peroxide (H2O2). The liquids are separated by a single oxygen atom.
Once the digital image is acquired, image processing science takes over by segmenting objects of interest, iteratively transforming the pixels, extracting digital features and mapping them against what the software believes is the digital signature of an object. We call this approach ’Signature Mapping™’. The software can explore as many as 1 million features, though for TB less than 100 are considered. Once prospective false positive objects are removed the software classifies the case as positive, assigning it a load factor, or it is classified as negative.
Using embedded technology that is already present in the laboratory and familiar to the technician is one of the strengths of TBDx™:
- The microscope is used as a digital input device for the image analysis technology.
- Human vision is supplemented when a camera is added.
- Software makes it possible to automatically focus the image and to capture one or hundreds of digital images, often in fractions of a second, completing the entire process in five minutes or less
- Software makes it possible to acquire enough images to classify a case as P+++, or acquire an extensive number of images to find the elusive TB in a scanty slide.
- All of the images are stored for further analysis.
How well does the software perform? The recent testing conducted at the National TB Reference Laboratory in South Africa will soon be publishable and provide an answer to that question based upon an independently conducted evaluation of the technology. At a preliminary and summary level, TBDx™ appears to have performed very similar to our internal testing and our expected performance outcomes. In the meantime, one place to turn for insights can be found in a November, 2012 blog post describing a poster presentation by Dr. Dave Clark, in Kuala Lumpur.
From all of our internal tests we have seen sensitivity range as high as 89%. However, this sensitivity declines as adjustments are made in the algorithm to remove false positive cases and improve specificity. And here is where we see a terrific opportunity to combine the strengths of TBDx™, a fully developed and very sensitive computer-vision technology, with a molecular test such as GeneXpert, that is sensitive, but also very specific with few false positive cases.
We will explore the combinations of the two technologies and their benefits in the final installment of this three-part series.