One scientist’s quest to teach machines to see cancer the way doctors do

The conceptual and technological underpinnings of digital pathology have been around for decades, with the virtual microscope emerging in the 1990s. But mainstream adoption remained elusive. For instance, when the digital histopathology service company HistoWiz launched in 2013, its focus was straightforward: provide faster turnaround times than traditional histology labs. Over the years, the company expanded its services while developing PathologyMap, a digital platform for managing and analyzing histopathology slide images that would grow to house one of the world’s largest collections of preclinical data.

Then came the pandemic. As social-distancing restrictions forced pathologists away from their microscopes, the field reached a turning point.

“Let me provide some context,” said Akash Parvatikar, Ph.D. “HistoWiz secured $32 million in funding during the pandemic. We were one of the only labs that remained operational in New York during that period,” said Parvatikar, the company’s product manager and AI scientist, who joined in 2022. The company was also involved in advancing vaccine discoveries during the pandemic, which helped drive a 2021 funding round.

The Series A financing, led by Vivo Capital with participation from venBio, Asahi Kasei, and Jon Oringer, signaled the industry’s recognition of digital pathology’s role in drug development.

While COVID served as a kickstart, the industry’s focus has since shifted toward more gradual, long-term adoption. “While significant investment went into the clinical space initially, we’re now observing a shift toward the preclinical space based on recent industry developments,” Parvatikar said.

In recent years, the number of alliances between digital pathology companies has expanded significantly—from August 2022 to the same month in 2024—according to research from DeciBio. Reflecting growing interest from hospitals, research institutions, and commercial labs, DeciBio found that almost 50% of the partnerships in this field were forged over the two-year period.

The industry is currently in a phase of “soft adoption” rather than aggressive implementation, Parvatikar noted. While AI momentum continues to advance swiftly, a degree of conservatism in highly regulated fields such as pathology and drug discovery limits the scope of adoption. “Pharmaceutical companies often focus on basic but high-impact capabilities,” Parvatikar said. For example, they might pursue AI-based digital pathology tools for mitosis detection or basic cell segmentation and quantification within specific regions. “These seemingly simple tools can save many hours of manual work that was previously required,” he said.

Yet proposals for more aggressive digital pathology deployments can attract skepticism. “We often hear the question: ‘Why should we digitize these slides?'” Parvatikar said. “We’re already paying real estate costs for physical storage, so why invest millions in digitization and cloud storage?”

Image similarity search as a ‘killer app’ for digital pathology

Every new technology platform needs a proverbial “killer app”—a technology so transformative that it redefines the entire market. From VisiCalc driving Apple II sales in 1979 to Lotus 1-2-3 cementing the IBM PC’s success to the debut of the iPhone, history shows that widespread adoption often hinges on a novel technology hitting the market at the right time. For digital pathology, that moment of transformation came unexpectedly through a global crisis: the COVID-19 pandemic forced a rapid shift to remote work, suddenly making digital slides and virtual collaboration essential rather than optional for a significant number of labs.

Parvatikar identifies image similarity search as a killer app for digital pathology. “This is both specific and broad in scope,” Parvatikar explains. Image similarity can be based on metadata or specific histological features. For instance, if a laboratory has a repository of 10,000 digitized slides and receives a challenging new specimen, image similarity search allows pathologists to swiftly identify and retrieve similar cases from the database.

This capability could significantly advance both drug discovery and clinical diagnostics, Parvatikar added. “It enables rapid comparison of patient histories and similar cases, allowing pathologists to make more informed decisions,” he said. “While it’s technically possible to do this with physical glass slides, the volume and speed at which you can perform these comparisons are severely limited with analog methods.”

In drug discovery, similarity search enables researchers to quickly identify and eliminate unsuccessful approaches early in the development process, significantly reducing costs. “This is where AI and digitalization will be transformative. While we’ll still need scientific expertise to identify promising compounds and targets, AI simulation capabilities can help eliminate thousands of potentially unsuccessful trials before they begin. With today’s computing power and AI capabilities, we can dramatically reduce the resources spent on approaches unlikely to succeed,” Parvatikar explained.

Getting AI to think like a pathologist

But implementing AI effectively in digital pathology requires more than just raw computational power—it demands an understanding of how medical experts actually work and think. The challenge of early cancer detection has always been as much about human cognition as it is about technology. Consequently, while studying at Carnegie Mellon University and the University of Pittsburgh School of Medicine, Parvatikar did more than study technology and biology when earning his doctorate in computational biology.

“During my Ph.D., I focused on understanding how pathologists make diagnostic decisions,” Parvatikar said. “I took cognitive psychology classes to better understand the diagnostic process. The goal was to develop AI systems that could support pathologists while maintaining transparency in their decision-making.” He built an explainable AI framework for detecting breast cancer, which was “very close to talking the language of a pathologist so they could understand what the AI is doing,” he said.

For pathologists, knowing where not to look is just as crucial as knowing where to focus their attention. A single glass slide image contains an immense amount of visual data—equivalent in size to multiple tennis courts when printed. “You’re essentially looking for subtle indicators that could signal early-stage cancer,” Parvatikar notes. This makes early detection challenging, especially in pre-invasive cases where the signs are subtle and interpretations can vary among pathologists.

Parvatikar’s approach also drew inspiration from scene recognition research, widely used in applications of human-computer interaction. “Just as we identify a kitchen by recognizing objects like refrigerators and stoves and their spatial arrangement, we applied similar concepts to tissue analysis,” he explains. Unsurprisingly, cancer diagnosis proved more complex than identifying household objects. “In cancer diagnosis, especially pre-invasive breast cancer, the features are more amorphous compared to scene recognition. There’s no strict statistical rule that says, ‘If x percentage of this feature is present, it must be diagnosed as y.'”

This ambiguity became an opportunity for innovation. “We developed an unsupervised learning approach to understand how these diagnostic features interact,” said Parvatikar. The team focused on cases where pathologists had different interpretations, using these challenging examples to build more robust AI systems.

The goal wasn’t to replace human expertise but to enhance it. “Our AI model highlighted the most diagnostically relevant regions within the image where physicians should focus their attention,” Parvatikar emphasized. “While some cases show obvious signs, particularly in advanced stages like metastatic breast cancer, pre-invasive cases require much more subtle analysis.”

This emphasis on transparency is critical for clinical adoption. Doctors,  just like other researchers, want to see the data supporting a new finding. “AI is often thought of as a black box,” Parvatikar noted. “But for medical diagnosis and FDA approval, you can’t just deliver 95% accuracy without explaining the failure modes and decision process.” To bridge this gap, Parvatikar built an explainable AI framework that could “speak the language of pathologists.” The system would identify specific regions of interest and articulate its reasoning: “These are the regions I’m examining, this is what I’m detecting, and here’s how I’m combining these features to reach a diagnosis.” The collaborators at UPMC and Magee-Womens hospital team then validated this approach by comparing it with pathologists’ diagnostic processes.

“By highlighting relevant regions and providing clear reasoning for its suggestions, the AI helps bridge the gap between computational analysis and human expertise,” Parvatikar concludes. “The goal is to build trust through transparency, especially for critical medical decisions where understanding the ‘why’ is just as important as the diagnosis itself.”

Laying the groundwork for the future of digital pathology

While AI hype has certainly crescendoed in recent years, more early adopters are finding value in strategic deployments. “The key is understanding that AI isn’t simply a button you press; it requires careful implementation and understanding of the underlying technology,” Parvatikar said.

With that in mind, HistoWiz aims to demonstrate concrete value to stakeholders. “We need to effectively communicate that digitized slides enable cloud access, seamless sharing capabilities, and the ability to run AI tools that accelerate research.”

The approach is paying off. The company recently commercialized its PathologyMap 2.0 platform, its latest digital slide management system for histopathology research. It offers features such as high-speed search, metadata tagging, and on-platform AI analysis. “We’ve received excellent feedback,” Parvatikar said. “In just two months, we’ve onboarded over 100 customers, including major pharmaceutical, biotech, and academic institutions.”

The platform’s success stems from its comprehensive approach to integration. “We’re partnering with several AI companies to integrate various capabilities into our platform, including detection of necrosis, tumor stroma, and cell segmentation,” Parvatikar explained. The company’s goal is to “make it easier for researchers to interact with AI tools and bring multiple partners together in one place.” This integration strategy enables sophisticated functionality: “This allows our customers to perform multi-AI analysis on a single platform. For example, they can run tumor detection from one AI provider and necrosis detection from another, all through a unified interface on PathologyMap.”

HistoWiz’s customer roster includes more than 500 organizations, including Sanofi, Regeneron, Bristol Myers Squibb, Johnson & Johnson, Calico, Revolution Medicines, and L’Oréal.

The growing traction comes at a time when there are more signals that the pharma sector’s R&D efficiency may be unsustainable. A recent Harvard and Johnson & Johnson study published in Nature revealed that drug development costs are approximately double previous estimates—around $5.5 billion per successful drug when accounting for both successes and failures across thousands of companies. This figure is far higher than earlier estimates of in the ballpark of $2.3 billion that focused only on major pharmaceutical companies.

“While identifying the right therapeutic targets is important, I believe there’s more immediate value in being able to efficiently rule out unsuccessful approaches early,” Parvatikar said. “This is where AI and digitalization will be transformative,” Parvatikar said. While there is still need for scientific expertise to identify promising compounds and targets, AI simulation capabilities can help eliminate thousands of potentially unsuccessful trials before they begin. “With today’s computing power and AI capabilities, we can dramatically reduce the resources spent on approaches unlikely to succeed,” Parvatikar concluded.

---

Filed Under: machine learning and AI, Oncology
Tagged With: AI-assisted diagnostics, Akash Parvatikar, digital pathology automation, drug development costs, histopathology digitization, HistoWiz, image similarity search, PathologyMap platform, preclinical imaging
 

---