Deep Neural Networks in image processing, predictive modelling and diagnostic decisions
Posted 2nd September 2020 by Liv Sewell
Now more than ever intelligent human beings are needed to fulfil high value and high complexity tasks and histopathology is no exception. In fact, it is one of the sectors in the most desperate need.
Correct interpretation of the complex and subtle patterns in a human tissue sample is a task that requires years of training to develop the level of skill and experience necessary.
The supply of histopathologists with the required level of skill and experience cannot keep pace with the burden of demand for tissue assessments and the situation is only going to worsen as the population increases and ages. Thus, alternative solutions are needed to reduce the burden on histopathologists and the expansion and growth of Information and Communications Technology offers an opportunity to automatically, objectively, rapidly and non-laboriously process digital images of human tissue samples.
Digital advances in histopathological examination
Great advances have been made in conventional digital imaging whereby extremely precise and accurate colour images can be taken of histopathological slides. These digital images have sufficient detail and clarity to act as a facsimile for a real histopathological slide in concert with a high magnification microscope and thus eliminate the need for the slide to be physically present during the histopathological examination.
This has been the first great stride in accelerating the processing of tissue samples as remote working has become possible and thus maximising the amount of time that can be spent per day assessing tissue samples. However, conventional digital imaging does not offer any new or complementary data source that could aid the diagnostic decisions.
The real frontier in medical imaging and diagnostics
The more recent advancements in optical physics such as Near Infra-Red imaging and Raman imaging offers the opportunity to collect new and complementary data sources by imaging at non-visible modalities to extract a chemical image of the tissue sample. These modalities also provide an element of depth into the tissue so the measurement goes beyond a plain surface image.
Such chemical images can provide molecular profiles at very fine granularity and thus provide a means of discriminating between normal and abnormal cells and hence can provide such a new and complimentary information to the diagnostic processes. Armed with such new and complementary information better diagnostic decisions are possible.
Historically the computing power required to process large histopathological images and large chemical images was beyond reach and similarly the artificial intelligence tools that were envisaged for rapid image processing and predictive modelling were unachievable.
In the last decade, there has been an explosion in computing power and capacity and these Deep Neural Networks have been brought into reality for a wide variety of tasks. In recent years, they have found tangible application in histopathological diagnoses and a growing branch of research has used Deep Neural Networks for conventional image processing and associated diagnosis.
Deep neural networks hold the potential to both innovate diagnostics and reduce the burden on the shortening supply of histopathologists
There is no reason to believe that Deep Neural Networks cannot also be used for processing chemical images and thus it is possible to fuse conventional and chemical histopathological images to form complementary image counterparts that can synergise to yield greater useful datasets for the Deep Networks to analyse.
These Deep Networks can literally learn from the best as they are provided with image labels from highly skilled and experience histopathologists. Thus, human expertise can be immortalised and be indefinitely replicated and thus the chronic shortage of histopathological expertise can be solved.
Dr. Patrick Jackman is a Senior Principal Scientist at the Technological University Dublin with expertise in the integration of the state of the art in hardware and software to improve workflows and system performance.
Leave a Reply