Newsbytes

The team also combed through all cases to find the most debated diagnoses. Many of these were easy to identify, Dr. Schaumberg says, because there tends to be a disproportionate amount of discussion on Twitter when there is disagreement. For such cases, the algorithm took a majority vote, and if this approach did not automatically infer the correct diagnoses, the team annotated the cases themselves rather than use the text from the associated tweets.

Another, perhaps surprising, step in creating the data set involved collecting social media content that had nothing to do with pathology, such as pathologists’ vacation pictures posted on Twitter. To develop a machine-learning classifier to identify histopathology stains, Dr. Schaumberg and colleagues had to include in the data set histopathology and nonpathology images, so they obtained consent to download images of ski slopes and birthday parties from some of those who contributed relevant data and manually separated these from the pathology-related tweets. The team plans to make their annotated data set of photomicrographs used in the project, sans personal photos, available under an open-source Creative Commons license by the end of the year.

Using this expanded data set, Dr. Schaumberg and his collaborators developed the supervised machine-learning and deep-learning models that can identify histopathology stains, discriminate between tissue types, and classify images by disease state. The project employs “the first pan-tissue pan-disease (i.e., from infection to malignancy) method for prediction and search on social media, and [is] the first pathology study prospectively tested in public on social media,” according to the Modern Pathology article.

Yet Twitter, like all social media platforms, has its drawbacks—among them, the difficulty of finding comparison cases for rare disease entities. It was this problem that inspired Dr. Schaumberg to incorporate data from PubMed into Pathobot’s data set.

Dr. Schaumberg recognized the need to extend Pathobot’s reach when Daliah Hafeez, MD, shared a liver subcapsular case on Twitter in which it was unclear whether the rounded structures shown in the photomicrographs were parasite eggs, nematodes, or vegetable matter mimicking disease. Dr. Stayerman responded “Perforation? Fistula?” to indicate how lentils may have reached the liver. “Pathbot found a similar lentils case from a colleague, and our internal tools for searching PubMed also found similar lentil cases,” Dr. Schaumberg says. Dr. Hafeez concluded that the patient had a duodenal fistula that had healed. “Vegetable matter it is,” she posted to Twitter.

It was after using internal tools to search PubMed for cases that might provide additional context for the lentil finding and posting those cases to the Twitter thread that Dr. Schaumberg realized Pathobot could perform an automated PubMed search for pathologists. To find relevant articles for the PubMed data set, the team used its machine-learning system trained to differentiate H&E images from all other images. The system filtered more than one million PubMed articles for figures that included images of H&E-stained tissue.

The number of cases in Pathobot’s social media database can also be a limiting factor as a result of too few similar cases being returned for comparison. To address this shortcoming, Dr. Schaumberg encourages pathologists to contribute data by sending a direct message to @pathobot on Twitter or going to his website, www.pathobotology.org/contact. He also encourages users to repeatedly use @pathobot on the same case as more photomicrographs or diagnoses emerge to help refine Pathobot’s search results.

As Dr. Schaumberg delves deeper into AI and labors over Pathobot, he’s looking to the future. His plans include incorporating new sources of data, such as The Cancer Genome Atlas, into the project and shoring up gaps in the current data. “Extending the data set to whole slide images may make our methods more applicable to hospitals since hospitals often have the resources to invest in whole slide imaging,” he says.

“It’s exciting that AI can be applied at the clinical level rather than just the basic science level,” he concludes. “To me, that’s very motivating.” —Charna Albert

X-Lab making inroads in laboratory network outreach

X-Lab has announced that it is focusing on bringing its product for digitizing and automating the laboratory test referral and results data-reporting process to the United States. The system, which will be marketed under the name Labgnostic, is designed to connect diagnostics laboratories worldwide through a single connection to an interoperable hub.

In March, the United Kingdom’s National Health Service mandated that all NHS trusts transfer externally referred SARS-CoV-2 test requests and results through the system, which is marketed under the name NPEx in the United Kingdom. The NHS selected the technology solution for its ability to share workload capacity information and provide faster turnaround times and safer data exchange.

“With a hub connecting diagnostic systems across the UK, Ireland, and into Europe, X-Lab [is] now exploring use cases for this technology across other continents,” according to a press statement from the company.

X-Lab, 424-367-8081

Dr. Aller practices clinical informatics in Southern California. He can be reached at raller@usc.edu.