How Pathware is accelerating pathology by delivering AI-powered expert analyses

Case Studies Aug 24, 2020

Pathware creates software that simplifies digital pathology workflows and reduces the rate of burdensome repeat procedures. Their flagship product, Bioptic is a cloud-based sample management platform that returns a biopsy quality assessment in just a matter of minutes - verifying that a high-quality sample was collected with greater than 90% certainty.

Pathware’s BIOPTIC™ sample management platform

Long before Covid-19, Pathware’s founders believed that distributed teams of pathologists can accurately assess the quality of tissue samples more cost-effectively and quickly. Their CEO, Michael Moore, and CTO, Jaron Nix, saw firsthand how wasteful it was for experienced pathologists to spend the majority of their job on routine tasks that were 90% visual and could be accelerated by the latest advances in computer vision. This especially rings true for assessing biopsies, a primary responsibility for pathologists. One in every five needle biopsies fail to collect a sufficient sample to make a cancer diagnosis, resulting in weeks of intermediary stress on patients and creating $3.6B in annual losses, according to Cytojournal.

Saving time when saving lives

The primary use case for Pathware centered on solving the problem of pathologists needing to spend hours on routine visual procedures. Prior to being able to provide a cancer diagnosis, pathologists traditionally spend roughly 1-hour assessing the quality of all tissue samples collected using traditional microscopy methods which are time-intensive and manual. Pathware’s product vastly reduces the time clinicians spend assessing a slide by providing the key tissue information and sample metrics, all calculated from a whole slide image that is annotated using Labelbox. With enhanced automation, patients can get their results faster which starts with knowing whether or not their biopsies contain enough tissue for a diagnostic result to be determined.

Pathology is an expensive and messy process. Pathware addresses two crucial parts of the process using computer vision by rapidly labeling tissue samples with bounding boxes (pictured above) to create high quality training data. These labels help classify assessments and answer the questions of: (1) What are we looking for in this sample and (2) can we use this sample based on its sufficiency?

In order to deliver the analyses in an accurate and predictable manner, the Pathware team needed to label large amounts of tissue sample assessments, and then feed the data into their machine learning models. Building a solution in-house was not an option given that the team did not have any web developers and user interface designers on staff, nor was training data infrastructure part of their core focus and expertise. Using Labelbox as the central hub for their training data saved their team months of R&D work by delivering an easy to use and cost-effective SaaS solution for annotating, managing, and iterating on training data.

To set up the images for labeling, Pathware created thousands of tissue clusters on a single pathology slide and uploaded them into Labelbox. These were presented one by one as a series of images that were classified with bounding boxes and image masks as either having no relevant features or relevant features. The goal for labelers was to look for a threshold of critical mass, and then adequacy was verified by third-party pathologists as true or false. To give an idea of how it looks, an entire pathology image in raw format could resemble a night sky. There was inherent complexity in generating the slide image but the Pathware team found that chopping it up into 5-12 x 5-12 images created an ideal situation to teach a machine to look for adequacy. Jaron’s team also found that this size and magnification level was just enough not to overwhelm a pathologist and would take roughly 10 seconds to label each image.

Deriving value from a training data platform

The primary pain points that Labelbox solved for Pathware was providing a way to avoid having to build and maintain costly infrastructure for managing their training data. Instead, their team of fifteen were able to immediately focus on the hardware and core application without worrying about building a training data platform from scratch or relying on open-source software that had incomplete features. The combination of Labelbox’s simple and intuitive user interface coupled with Pathware’s team of trained pathologists allowed expertise to be leveraged without delaying operations or sacrificing efficiency.

One specific feature which Pathware’s team utilized most was Labelbox’s ontology management features. This feature was especially valuable given they outsource much of their annotation work to external labelers. Companies earlier in their ML maturity tend to create massive ontologies that serve as exhaustive lists. The Pathware team learned that this was not necessary, given that speed and the ability to scale came directly from their decision to have a streamlined ontology.

“Labelbox gave us a lot of versatility by creating thoughtful ontologies with a streamlined interface. It was a critical feature and enabled us to produce high-quality labeled data with minimal errors and inconsistencies. In addition, we really liked the infinite flexibility of how classifications are set up and the product was intuitive for our pathologists to use.”  (Jaron Nix, Pathware CTO)

The future of diagnostic medicine

Pathware’s current work focuses on thyroid samples which they see as the most straightforward use case to tackle first. Future projects will involve more nuanced solutions such as detecting breast and cervical tissue samples. Pathware plans on continuing to refine its product and to have over-the-air (OTA) updates, similar to Tesla’s electric cars which combine hardware and real-time software. The near-term goals are to expand on their labeled datasets and to work closely with the FDA in order to release updates when a new verification data is ready.

Labelbox Benefits:

  • Easy-to-use and cost-effective SaaS solution for training data vs. building internally
  • A faster and better experience for end-users (pathologists)
  • Increased confidence and reduced time going into ML production
  • A centralized location and hub for all of Pathware’s training data
  • Flexibility and ability to scale to other future use cases


Labelbox is a collaborative training data platform empowering teams to rapidly build artificial intelligence applications.