StreamingCLAM: Scalable and collaborative AI starts here

How do you train AI models without starting from scratch every time new data becomes available? That challenge is central to the future of AI in pathology, and it’s what inspired StreamingCLAM, an end-to-end learning tool designed to update AI models with global context, instead of using only patches.

Developed by Bigpicture member Stephan Dooper, PhD candidate at Radboudumc, StreamingCLAM is a end-to-end learning tool shaped by the project’s ambition to make AI development in pathology more scalable, sustainable, and collaborative. While not a formal Bigpicture deliverable, it reflects the same values — and is ready to support the platform’s long-term goals as its data infrastructure grows.

“We wanted to move towards more generic, pan-cancer algorithms that don’t rely heavily on manual annotations,” Stephan explains. “Such a model is hard to build, especially when you’re working with small, organ-specific datasets. The idea behind StreamingCLAM is to learn with fewer labels and to be organ-agnostic.”

What is StreamingCLAM?

StreamingCLAM combines two AI concepts into one framework: CLAM (Cluster-Attention Multiple instance learning) and streaming.

Originally, CLAM was developed for weakly supervised learning, allowing AI models to learn from slide-level labels by identifying which regions in whole slide images (WSIs) are most informative, without the need for detailed annotations. It uses an attention-based mechanism to focus on the most relevant image patches, making it particularly useful in pathology, where annotations are costly and time-consuming.

StreamingCLAM takes this architecture a step further by integrating a streaming approach. This enables the model to update, using the entire global context of the tissue, whereas normally AI tools only learn using smaller tissue patches, which are limited to local context only. “The result is a model that looks at the bigger picture instead of individual puzzle pieces,” Stephan explains. “This makes the tool especially valuable for practical, collaborative settings like the platform Bigpicture is building.”

A sneak peek at StreamingCLAM in action

A future-fit tool for Bigpicture users

As Bigpicture’s infrastructure continues to grow, with more harmonized datasets and shared features becoming available, StreamingCLAM is ready to support the platform’s users. Researchers and developers could eventually use it to test cross-cohort model performance, update algorithms on the fly as new data comes in or run large-scale weakly supervised studies. “The kind of infrastructure Bigpicture is building is exactly what’s needed to train and update models like StreamingCLAM at scale,” says Stephan.

What’s still being improved?

StreamingCLAM is not ‘finished’. One of the key areas of active development is ensuring the tool can handle uncertainty in real-world data and predictions. “Right now, I’m working on improving how StreamingCLAM quantifies and manages uncertainty across different types of inputs,” Stephan says. “This is essential if we want it to support robust decision-making in clinical or cross-institutional settings.”

These improvements are designed to support future integration into shared environments, making it easier for users across Europe (and beyond) to plug StreamingCLAM into their workflows as the Bigpicture platform evolves.

AI in a fast-moving landscape

The pace of change in AI has accelerated dramatically in recent years, reshaping both research and application. “Four years ago, foundational models were just starting to emerge, now they’re everywhere,” says Stephan. “You’re constantly adapting, but that also means you’re always learning.”

That pace of change can be challenging, especially in a PhD trajectory. But for Stephan, it’s energizing, particularly in a field where the stakes are high and the impact is real. “There’s so much work left to do in medical AI. I see this evolution not as a burden, but as an opportunity. It means we’re moving forward.”

As Bigpicture is evolving from platform development to activation, tools like StreamingCLAM could empower users to build models collaboratively, improve diagnostic tools faster, and ultimately bring benefits to patients across Europe and beyond.