Dr. Clark’s vision was to improve sputum microscopy by eliminating human factors (areas for potential misdiagnosis and errors) from the diagnosis of TB. However, his vision of automated sputum microscopy is not a new concept. Over many years various attempts at automated sputum microscopy have occurred, most specifically from academia, through experimentation with algorithm-based approaches for the detection of TB bacilli on a single microscopic slide, as compared to the two slides per patient in the current manual process. This university-based research has clearly demonstrated that algorithms can be used to automatically detect and classify objects as TB bacilli. However, the research was fundamentally flawed due to the narrowness of the dataset. Among the weaknesses of their approach were the following: (i) the number of research and test cases were very limited and not representative of the universe of potential cases, (ii) the digital images acquired were not captured from an automated focus process, but rather through manual focus by the microscopist, and (iii) staining quality was carefully selected to include only the highest quality specimens.
Most of the university-based researchers have published articles on their work with automation of smear microscopy, highlighting the limitations of their approach, yet reaching the conclusion that while automation may be extremely challenging, there are very good reasons for pursuing it.
Automating the smear microscopy process to deliver consistent and accurate diagnostic results is no trivial matter. There are challenges and essential barriers that must be overcome.
To expand upon the research already performed required a more extensive research approach. We needed to further understand the current smear microscopy laboratory protocols, as well as the areas where the manual process of TB diagnosis falters. Furthermore, we needed to understand the demand trajectory for smear microscopy and its economic landscape. As we moved the early proof-of-concept version of TBDx forward, it became readily apparent that we would need to provide end-to-end automation of the process post-staining of the slides…the human injected too many variables that affected a consistent result. Our R&D plan included the following requirements:
- Assemblage of a hardware platform capable of high throughput with minimal human involvement
- High quality image capture for both automated detection, but also human interpretation, if desired
- Automation of data and reporting to eliminate redundancy and human recordation errors
- Intelligent workflow processing to provide expanded diagnostic flexibility
- Ability to distribute images and reports to a “cloud service”
- Robust reporting and communication systems – ability to automatically generate results to multiple recipients via text messaging, email or voicemail
We spent much of our early research and development efforts in locating, reviewing and testing the various components of the hardware platform: autoloaders, barcode readers, auto-stages, digital cameras and computer monitors. When the selection of components had been completed the real work started. We had to develop and write the computer automation programs to seamlessly interface the operations of each component into an integrated hardware solution.
Today we have seamlessly integrated all the hardware components creating a highly flexible technology platform that enables the user to quickly and effortlessly acquire digital images from 1-200 slides with the touch of a single key or screen icon. Once initiated no human interaction is necessary until the process completes. Each slide, from the time the system initiates until it has completed processing, takes approximately two minutes.