NEW YORK – German artificial intelligence-based software company Mindpeak has developed a proprietary machine learning method to fully automate the process of digital pathology that can also be integrated with other laboratory software to help clinical laboratories streamline their processes.
As its base technology, Hamburg-based Mindpeak uses a proprietary hybrid deep learning method and combines that with its AI pre-trained with pathology data that doesn't have annotations from pathologists, founder and CEO Felix Faber said. It can be costly to train an AI algorithm entirely on annotated, or "supervised," data, since it requires pathologists to review the slides and manually annotate them, so starting with images without annotations gives the algorithms a baseline to build from, Faber said.
Data that includes annotations from pathologists was inputted into the algorithms after the pre-training.
Using multiple kinds of data has made the resulting algorithms "robust against variations" that can occur in normal pathology labs, Faber said. Slides can vary based on the staining machine used, the scanner, the temperature in the lab, and other environmental elements, and the algorithms are able to account for that variation and "can handle it very well," he noted.
Originally, Mindpeak's tech was developed to meet pathologists' need to quantify images of immunohistochemistry-stained slides taken from microscope cameras, Faber said. That was difficult as the algorithms had to account for differences in the types of microscope camera that could affect the quality of the image in addition to a lab's environment, he said, adding as digital pathology has shifted toward whole-slide imaging, the algorithms have been just as accurate.
As soon as an IHC-stained slide comes out of the wet lab and goes into a scanner, it is uploaded to the AI server, which will scan the image and detect every cell, determining whether it is a tumor or non-tumor cell, Faber said. The AI processing occurs in parallel with the scanning of a batch of slide images, so once one slide is scanned and uploaded, the AI will process it while another image is scanned, leading to almost immediate results approximately two to five minutes after all of the images have been scanned.
Additionally, in the company's newest versions of its algorithms, there's no manual intervention required, beyond a manual check by a pathologist at the end of the process, according to Faber.
Mindpeak isn't the only company to offer AI-based digital pathology — firms like PathAI, Paige, Visiopharm, and Aiforia have similar products — but Faber said that Mindpeak's algorithms need fewer images and fewer annotations to be trained than most other AI algorithms.
While most of its algorithms currently only work with IHC-stained slides, the company is developing algorithms for uses with other stains, Faber said, such as its Onychomycosis AI algorithm, which detects nail fungus hypha from periodic acid-Schiff-stained slides. The onychomycosis dermatopathology algorithm is currently available in the US as a research-use-only product.
Although the algorithms can be purchased on their own and integrated into a laboratory's existing workflow software, Mindpeak has also integrated its platform with six vendors across the US and Europe, and its technology is used in 12 labs.
In one example, Mindpeak has combined its BreastIHC AI algorithm with Gestalt Diagnostics' lab automation software to create a way for pathologists to completely automate their workflows when dealing with breast tissue samples.
Gestalt's digital pathology software is integrated with a whole-slide scanner and is intended to "manage the image throughout its life cycle," said Lisa-Jean Clifford, Gestalt's chief operating officer and chief strategy officer. It's intended to be the "pathologist's cockpit" to handle the entire digital pathology workflow, she added, and can be customized based on a lab's needs.
Opko subsidiary BioReference Laboratories currently uses the Gestalt and Mindpeak combination, switching to it from a platform that was being discontinued, said Ellen Beausang, senior VP of advanced diagnostics at BioReference. The reference lab looked for an upgrade to what it had been used to and found that the Gestalt-Mindpeak combo gave BioReference "the ability to have more innovative technology and AI resources to build a more robust offering," she said.
BioReference was looking for a system that was reliable, Beausang added, since the lab previously had used technology that was "unstable." It also looked for "something that could improve the turnaround time for results," as well as a platform that offered enhancements that included AI. She added that "the ability to enhance the offering with new technology additions and/or the ability to add new testing areas was important."
As for the pricing of these offerings, Mindpeak uses a per-case business model, Faber said — "you really only pay for what you need." The price is dependent on the volume of slides being analyzed and how many times per day an algorithm is being used, he said.
The BreastIHC algorithm is CE-IVD marked in Europe and marketed as a laboratory-developed test in the US, with Mindpeak aiming to get clearance from the US Food and Drug Administration in 2023, Faber said, although the team is still discussing whether it will pursue 510(k) or de novo clearance, he added. It is also fully prepared for Europe's upcoming In Vitro Diagnostic Regulation, set to take effect in May, Faber noted.
The algorithm currently isn't reimbursed by insurers, but Faber said the firm plans to look further into reimbursement once it's approved by the FDA.
He added that the company is looking to expand to more distribution partners, with an emphasis on moving into China next.
The firm also has other products in development, including its PD-L1 Quantifier algorithm, intended to detect and quantify cells in non-small cell lung carcinoma from IHC-stained slides, Faber said. In addition, the firm has Ki-67 and estrogen/progesterone algorithms that are both CE-marked.