Slides in the Machine
Digital pathology is the future of storing and sharing images of tissue – and combining it with deep learning could further transform the field
David West, Jr. and Hunter Jackson |
At a Glance
- In the past, digital pathology has been held back by software and storage issues – but advances such as cloud computing are breaking down these barriers
- Whole slide imaging can improve efficiency, increase automation, and improve collaboration between institutions all over the world
- Machine-learning-driven “computational pathology” may also result in new biomarkers, workflow improvements, and image-based diagnostics, providing pathologists with new tools for diagnosing disease
- FDA approval of the first WSI system is a turning point for the technology, and soon “digital pathology” may instead simply be “pathology”
Artificial intelligence in healthcare is gaining ground in areas ranging from patient care to diagnosis, data management, and many others. And although we’re still a long way from reaching full automation, pathology is ripe for major disruption.
Gone are the days of having cabinets full of slides that are difficult to share and only viewable by squinting down a microscope. Pathologists can now examine tissue on their computers from anywhere in the world via whole slide imaging – and, with the help of computational pathology software powered by intelligent machine learning, will have new, precise information at their fingertips. The digital revolution is already underway.
From slides to files
Today, digital whole slide imaging (WSI) allows the capture of the entire tissue sample on a slide, as opposed to using a microscope camera attachment to capture a field of view. A modern scanner takes between 30 seconds and two minutes to capture a slide, usually at a magnification of 20x or 40x. The resulting image files are very large, allowing us to capture a great deal of data, but typically require specialist software to view and sophisticated IT solutions for storage. It’s not an incredibly new system (it has existed for research and education uses for some time) but we can expect to see increased adoption of WSI as technology advances and as storage issues are addressed. Progress will result in a new realm of possibilities for tools that enhance how pathologists review cases.
On efficiency and algorithms
Simply switching to a digital workflow has been shown to increase efficiency by 10 to 15 percent (1). But beyond these baseline improvements lie exciting and transformational benefits. Once we’re in the “digital realm” and dealing with pixels rather than glass, we open up two key possibilities. Firstly, distance no longer matters. Sending slides between two institutions on different sides of the world becomes much easier, opening up new telepathology possibilities, helping pathology departments and private labs to grow, and improving access to subspecialty experts for patients in remote areas.
Secondly, image analysis algorithms will be able to operate on the images. These are already used for automated or semi-automated immunohistochemistry quantification, helping to drive standardization and speed of analysis – and new methods are in development to augment H&E. The area could play a profound role in reducing inter-observer variability for many cancers, and in driving faster, more precise quantitative workflows.
There’s certainly some resistance to going digital, but most are rooted in challenges that are easy to overcome. One of those challenges is resistance to cultural change. We find the best way to address resistance is to give pathologists the opportunity to try it firsthand. Seeing is believing. It’s common for a pathologist to have negative preconceived notions based on experience with older digital pathology systems or just the “idea” of digital in general, so it’s especially rewarding to see a pathologist’s eyes light up when they see the quality of images, what quantitative tools can do today, and how simple collaboration with colleagues becomes. These systems can increase the efficiency of pathologists by automating monotonous tasks, such as identifying mitotic cells or screening benign tissue and identifying “cancer hotspots.” Pathologists can then focus their time on making informed diagnoses.
Another concern is image quality and throughput – and a year or two ago, this was indeed a larger concern than it is now. But these issues are improving every year, digital storage costs are going down, and the first WSI system was just approved by the FDA for clinical use – we are truly at a turning point.
As mentioned above, another great advantage to digital slide imaging is the potential for automated analysis. Deep learning is a class of machine learning algorithms that have been around for decades, with some major breakthroughs in the last few years. Newer learning techniques can be used to allow image and pattern recognition. The results so far have been remarkable (2), and are generating excitement in many fields, especially medical imaging.
There is now a well-timed confluence of events; the digitization of pathology so far is already generating lots of data, new methods are becoming available to technologists, and enormous amounts of computing resources are being made available via the cloud at low cost. All of this means that we can start training intelligent algorithms to recognize broad or specific patterns on a whole slide image, and translate features evident in the tissue into prediction (such as metastasis and recurrence) and classification (staging, grading, and differential diagnosis). This enables the creation of predictive biomarkers based on precise measurements of histological patterns, providing pathologists with new tools to answer questions about a given patient’s disease. This is especially useful where molecular tests fall short – image-based assays could be part of a portfolio of tests in a pathologist’s precision medicine arsenal. Scanned slides can be used to train deep neural networks to learn how the cellular morphology reveals genetic and epigenetic changes in the tissue. In some ways, it’s still early days, but this is already starting to be used in cutting-edge laboratories, and deep learning is likely to have a broad range of applications within pathology.
For example, cancer is a system and should be evaluated on a spectrum, so pathologists need many tools to uncover both molecular and morphological changes in the tissue to place a specific case of cancer on that spectrum. Many genetic tests can help genotype a patient’s cancer, which can help inform therapeutic options. However, to qualify for certain genetic tests, you must meet rigorous requirements, including overexpression of certain proteins and evident genetic alterations. With the advent of image-based diagnostics, it may be possible to expand the patient population that is eligible for these tests, informing precise treatment plans.
Is the future in the computer lab?
FDA approval of the first WSI system is exciting news, especially as many people thought it might never come. More vendors will likely follow, and in a few years, “digital pathology” may simply be “pathology.” It’s not going to happen overnight, but quantitative, computational-driven workflows will be one of the key drivers for market adoption, and laboratories that resist this change will likely fall behind more forward-thinking peers. As more slides are scanned and more data becomes available, we’ll likely see algorithms seep into the workflows of pathology labs. Early adopters are already using these algorithms to augment their work. In the next ten years, we anticipate that much of the professional component of pathology will be divorced from the physical laboratory, with human pathologists working in software driven “labs.”
- J Ho et al., “Can digital pathology result in cost savings? A financial projection for digital pathology implementation at a large integrated health care organization”, J Pathol Inform, 5, 33 (2014). PMID: 25250191.
- R Chen et al., “Identifying metastases in sentinel lymph nodes with deep convolutional neural networks” (2016). Available at: arxiv.org/abs/1608.01658. Accessed June 19, 2017.