Computer-assisted analysis using Google’s LYNA algorithm shows significant gains in speed required to analyze stained lymph node slides and sensitivity of micrometastases detection in two recent studies
Anatomic pathologists understand the complexities of reviewing slides and samples for signs of cancer’s spread. Two studies involving a new artificial intelligence (AI) algorithm from Google (NASDAQ:GOOGL) claim their “deep learning” LYmph Node Assistant (LYNA) provides increases to both the speed at which pathologists can analyze slides and improved accuracy in detecting metastatic breast cancer within the slide samples used for the studies.
Google’s first study was published in the Archives of Pathology and Laboratory Medicine and investigated the accuracy of the algorithm using digital pathology slides. Google’s second study, published in The American Journal of Surgical Pathology, looked at how pathologists might harness the algorithm to improve workflows and use the tool in a clinical setting.
Medical laboratories and other diagnostics providers are already familiar with the improvement potential of automation and other technology-based approaches to diagnosis and analysis. Google’s LYNA is an example of how AI and machine learning improvements can serve as a supplement to—not a replacement for—the skills of experts at pathology groups and clinical laboratories.
Early research done by Google indicates that integrating LYNA into existing workflows could allow pathologists to spend less time analyzing slides for minute details. Instead, they could focus on other more challenging tasks while the AI analyzes gigapixels worth of slide data to highlight regions of concern in slides and samples for deeper manual inspection.
LYNA Achieves 99% Accuracy in Study of Metastatic Breast Cancer Detection
According to the research cited in a Google AI Blog post, roughly 25% of metastatic lymph node staging classifications would change if subjected to a second pathologic review. They further note that when faced with time constraints, detection sensitivity for small metastases on individual slides can be as low as 38%.
In findings published in Archives of Pathology and Laboratory Medicine, Google researchers analyzed whole slide images from hematoxylin-eosin-stained lymph nodes for 399 patients sourced from the Camelyon16 challenge dataset. Of those slides, researchers used 270 to train LYNA and the remaining 129 for analysis. They then compared the LYNA findings to those of an independent lab using a different scanner.
“LYNA achieved a slide-level area under the receiver operating characteristic (AUC) of 99% and a tumor-level sensitivity of 91% at one false positive per patient on the Camelyon16 evaluation dataset,” the researchers stated. “We also identified [two] ‘normal’ slides that contained micrometastases.”
Google’s algorithm later received an AUC of 99.6% on a secondary dataset.
“Artificial intelligence algorithms can exhaustively evaluate every tissue patch on a slide, achieving higher tumor-level sensitivity than, and comparable slide-level performance to, pathologists,” the researchers continued. “These techniques may improve the pathologist’s productivity and reduce the number of false negatives associated with morphologic detection of tumor cells.”
Left: sample view of a slide containing lymph nodes, with multiple artifacts: the dark zone on the left is an air bubble, the white streaks are cutting artifacts, the red hue across some regions are hemorrhagic (containing blood), the tissue is necrotic (decaying), and the processing quality was poor. Right: LYNA identifies the tumor region in the center (red), and correctly classifies the surrounding artifact-laden regions as non-tumor (blue). (Image and caption copyright: Google AI Blog.)
Faster Analysis through Software Assistance
Rapid diagnosis helps improve cancer outcomes. Yet, manually reviewing and analyzing complex digital slides is time-consuming. Time constraints might also lead to false negatives due to micrometastases or small suspicious regions that slip by pathologists undetected.
The Google research team of the study published in The American Journal of Surgical Pathology sought to gauge the impact LYNA might have on the histopathologic review of lymph nodes for trained pathologists. In their multi-reader multi-case study, researchers analyzed differences in both sensitivity of detecting micrometastases and the average review time per image using both computer-aided detection and unassisted detection for six pathologists across 70 slides.
Using the LYNA algorithm to identify and outline regions likely to contain tumors, the researchers found that sensitivity increased from 83% to 91%. The time to review slides also saw a significant reduction from 116 seconds in the unassisted mode to 61 seconds in the assisted mode—a time savings of roughly 47%.
“Although some pathologists in the unassisted mode were less sensitive than LYNA,” the researchers stated, “all pathologists performed better than the algorithm alone in regard to both sensitivity and specificity when reviewing images with assistance.”
The Future of Digital Pathology using LYNA
While the two studies show positive results, both studies also reveal shortcomings. Google highlighted both limited dataset sizes and simulated diagnostic workflows as potential concerns and areas on which to focus future studies.
Still, Google’s researchers believe that algorithms such as LYNA will help to power the future of diagnostics as healthcare in the digital era continues to mature. “We remain optimistic,” state the authors of the Google AI Blog post, “that carefully validated deep learning technologies and well-designed clinical tools can help improve both the accuracy and availability of pathologic diagnosis around the world.”
While other industries see risk in the growth of AI, both studies performed by researchers at Google show how computer-assisted workflows and machine learning could accentuate and bolster the skills of trained diagnosticians, such as anatomic pathologists and clinical laboratory technicians. By working to compensate for weak points in both human skill and computer reasoning, the outcome could be greater than either AI or humans can achieve separately.
Google Creates AI to Detect When Breast Cancer Spreads
Google Deep Learning Tool 99% Accurate at Breast Cancer Detection
Google’s AI Software Seeks to Detect Advanced Breast Cancer Better Than We Have Before
Google’s AI Is Better at Spotting Advanced Breast Cancer than Pathologists
Google AI Claims 99% Accuracy in Metastatic Breast Cancer Detection
Applying Deep Learning to Metastatic Breast Cancer Detection
Assisting Pathologists in Detecting Cancer with Deep Learning
Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women with Breast Cancer
Data generated by medical laboratories and diagnostic providers takes an increasing role in treatment and precision medicine and allows greater analysis of data and integration of data into the care process
Most anatomic pathologists recognize that the unstructured data that makes up most pathology reports also represents a barrier to more sophisticated use of the information in those pathology reports. One solution is for pathology groups to adopt synoptic reporting as a way to get a pathology report’s essential data into structured fields.
The healthcare marketplace recognizes the value of structured data. In 2012, venture capitalists funded a new company called Flatiron Health. Flatiron’s goal was to access the medical records of cancer patients specifically to extract the relevant—and generally unstructured—data and put it into a structured database. This structured database could then be used to support both research and clinical care for cancer patients.
How valuable is structured healthcare data? Just this February, Roche paid $1.9 billion to acquire Flatiron. At that point, Flatiron had assembled information about the health records of two million cancer patients.
But Roche (ROG.S), recognizing the value of data, was not done. In July, it entered into an agreement to pay $2.4 billion for the remaining shares of cancer-testing company Foundation Medicine that it did not own. Foundation Medicine sequences tumors and uses that genetic data to assist physicians in diagnosing cancer, making treatment decisions, and identifying cancer patients who qualify for specific clinical trials.
Anatomic pathologists play a central role in the diagnosis, treatment, and monitoring of cancer patients. It behooves the pathology profession to recognize that generating, storing, analyzing, and reporting the data generated from examinations of tumor biopsies is a critical success factor moving forward. Otherwise, other players and stakeholders will move past the pathology profession and stake their own claim to capturing, owning, and using that data to add value in patient care.
How Lack of Standards Impact Transfer of Patient Data
DATAMARK Inc., a business process outsourcing (BPO) company headquartered in El Paso, Texas, reports that analysts from Merrill Lynch, Gartner, and IBM estimate unstructured data comprises roughly 80% of the information in the average electronic medical record. This data could be the key to improving outcomes, tailoring precision medicine treatments, or early diagnosis of chronic diseases.
From narrative descriptions of biopsies to dictated entries surrounding preventative care appointments, these entries hold data that might have value but are difficult to collate, organize, or analyze using software or reporting tools.
To further complicate matters, each service provider in a patient’s chain of care might hold different standards or preferred methods for recording data.
“At this point, [standards] are not to a level that helps with the detailed clinical data that we need for the scientific questions we want to ask,” Nikhil Wagle, MD, Assistant Professor of Medicine, Dana-Farber Cancer Institute, Harvard Medical School, and Associate Member, Broad Institute, told the New York Times.
An oncologist at the Dana Farber Cancer Institute in Boston, Wagle and his colleagues are creating a database of metastatic breast cancer patients capable of linking medical records, treatments, and outcomes with their genetic backgrounds and the genetics of their tumors. Despite best efforts, they’ve only collected 450 records for 375 patients in 2.5 years.
Nikhil Wagle, MD (above), Assistant Professor of Medicine, Dana-Farber Cancer Institute, Harvard Medical School, and Associate Member, Broad Institute, is building databases that link patient outcomes and experiences with their EHRs. But sharing that information has proved problematic, he told the New York Times. “Patients are incredibly engaged and excited,” he said, “[But] right now there isn’t a good solution. Even though the patients are saying, ‘I have consented for you to obtain my medical records,’ there is no good way to get them.” (Photo copyright: Dana-Farber Cancer Institute.)
Additionally, once records are obtained, the information—sometimes spanning hundreds of faxed pages—must still be processed into data compatible with Dana-Farber’s database. And updating and maintaining the database requires a full-time staff of experts that must review the information and accurately enter it as required.
When critical concerns arise—such as a cancer diagnosis—information that could yield valuable clues about treatment options and improve outcomes might be held in any number of data silos in any number of formats.
This doesn’t account for the complexity of organizing such information for researchers who are developing new treatments, applying data to less targeted approaches, or dealing with privacy concerns between care providers.
Moving forward, those who can create and interact with data in a way that requires minimal human touch to make it suitable for analysis, further processing, or archiving, could communicate data more effectively and glean value from the growing trove of data silos created by laboratories around the world.
Big Pharma Making Big Bets on Structured Data
These are all the reasons why the recent moves by Roche show the importance and perceived value of structured medical records data as it takes an increasingly important role in precision medicine treatments and diagnosis.
With its acquisition of both Flatiron Health and Foundation Medicine, Roche has secured the ability to generate data, convert said data into a structured format to drive decisions, improve core data-related services, and promote the value of their offerings. This positions Roche to maximize the value of its data for internal use and marketing to researchers and other interested parties.
For clinical laboratories, pathology groups, and other diagnostics providers generating untold amounts of data daily, this highlights a critical opportunity to stay ahead of future trends and position themselves as valuable sources of information as healthcare data continues to play an essential role in modern healthcare.
New Cancer Treatments Lie Hidden under Mountains of Paperwork
Unstructured Data in Electronic Health Record Systems: Challenges and Solutions
Pharma Giant Roche Just Made a $2.4 Billion Bet on Cancer Data
Roche to Buy Flatiron Health for $1.9 Billion to Expand Cancer Care Portfolio
Why Drug Giant Roche’s $1.9 Billion Deal to Buy Data Startup Flatiron Health Matters
Roche Acquires the Outstanding Shares of Foundation Medicine for $2.4Bn
New Solutions for Unstructured Data May Help with Clinical Laboratory and Anatomic Pathology Data