In this interview with Donald Allen of the Life Sciences Product Management team, Indica Labs proves its continuous innovation with the inclusion of a cutting-edge foundation model in HALO AI.

Thanks for taking the time today, Donald, for this interview. What is your current role and what do you do here at Indica? 

My pleasure. I’m the HALO AI Product Manager here at Indica Labs and I guide product development of HALO AI. In my role, I get the opportunity to help build tools and workflows that meet our customers’ needs to rapidly build, train, and validate AI for image analysis. 

What is the AI Annotation Tool in HALO AI? 

The AI Annotation Tool allows customers to quickly annotate a wide variety of objects in HALO® – from small objects like nuclei to larger objects like glomeruli or B cell follicles. It is a simple point and click workflow that allows customers to rapidly develop training data for HALO AI. 

What foundational AI model underlies the AI Annotation Tool?  

The Segment Anything Model (SAM) from Meta AI is the foundational model that powers the HALO AI Annotation Tool. If you’re interested in learning more about the foundational AI model, check out this blog post from Meta AI as well as their arxiv publication. Thanks to the PyTorch structure of HALO AI’s backend, we were able to easily integrate it into the product. 

What did SAM allow us to do that would have been harder had we developed the tool from scratch? 

Training a model like SAM demands massive resources. During SAM’s development, Meta AI leveraged over one billion masks across 11 million images, requiring immense compute power, infrastructure, and human effort that can cost millions. Leveraging Meta’s Segment Anything Model allows us to efficiently focus our resources on what matters to our users. Rather than developing a similar tool in-house, we can focus on crafting a user experience that meets our customers’ needs. 

Was SAM trained for digital pathology? 

The SAM model was not specifically trained on digital pathology data. This is what makes it a foundation model for image segmentation. It was trained on a broad and diverse dataset, making it highly adaptable and already valuable for digital pathology tasks. Internally, we are exploring fine-tuning SAM on digital pathology images to further improve the performance of the AI Annotation Tool on images that matter to our customers. We hope to share updates on this in the near future. 

How did Indica improve the AI Annotation Tool for the 4.1 release? 

In HALO 4.1, we updated the AI Annotation Tool with a dedicated user interface and more transparent behavior. The inference window is now explicitly defined, making it clear which objects are selectable by the tool. 

We’ve also added several messaging improvements: 

• The inference window is outlined in red while the model is still processing. Once inference is complete, the red outline disappears, and the user is free to select the objects. 

• Additional status messages inform you when the model is still loading to the Graphics Processing Unit (GPU), helping set clearer expectations. 

Finally, we introduced native FL image support. This means the AI Annotation Tool can now leverage FL image adjustments—like histogram adjustments or channel toggling—so it can accurately annotate the objects you see in the viewer. What you see, you can select. 

Are there other foundation models that are being evaluated for inclusion into the HALO and HALO AI image analysis platform for the 4.2 release or beyond?  

We are actively evaluating other foundation models for potential integration into the HALO platform. Their inclusion will depend on whether we can identify clear, practical use cases that meaningfully leverage the technology and improve user experience. For example, SAM was released by Meta in the middle of an active HALO development cycle, but its potential to power an AI-powered annotation tool was clear. It was obvious that our users would benefit from an AI-powered annotation tool and thus its integration was an easy decision. 

How do you go about evaluating the performance of models and considering if they are appropriate to bring into the HALO and HALO AI platform? 

Our AI Research team collaborates closely with other groups across the company including AI Diagnostics, Pharma Services, and the HALO platform development teams to explore the potential of foundation models in addressing challenging image analysis scenarios and datasets. This allows us to assess whether these models can improve the overall image analysis experience. We also benchmark foundation models against current state-of-the-art methods within the platform. When a clear and valuable use case emerges, we move toward productization and platform integration to make the capability broadly accessible to our users. 

What advantages might these foundation models bring to HALO customers?  

This is a complex question, as many different types of foundation models are now emerging in the digital pathology space. Broadly speaking, the hope is that foundation models improve the HALO user experience by reducing the time and effort required to train useful models, while also improving the accuracy and robustness of the resulting classifiers. 

By leveraging models that have been pre-trained on large, diverse datasets, digital pathology users may benefit from more consistent and reproducible results, and greater adaptability across tissue types and imaging modalities. Ultimately, foundation models have the potential to speed up R&D and enable more accurate and robust image analysis. 

Donald, thanks again for your time today. It is such an exciting time to be a part of AI in digital pathology, and we appreciate you breaking down how the Segment Anything Model was incorporated into HALO AI and sharing your thoughts on the future of foundation models in digital pathology software.