At UCLA's Ozcan Lab in Los Angeles, researchers have cracked a problem that has long haunted cancer diagnostics: how to make AI-powered pathology both accurate and trustworthy enough for hospitals and clinics to actually rely on it.
The challenge is real. Digital pathology — using AI to analyze tissue samples and spot cancer biomarkers — promises to transform how doctors diagnose and treat breast cancer. But most existing systems depend on expensive microscopy equipment and give doctors no way to know when an AI prediction might be wrong. That uncertainty can be dangerous. Prof. Aydogan Ozcan and his team set out to solve both problems at once.
Their innovation combines two technologies that don't usually work together: lensfree holographic imaging and deep learning with Bayesian uncertainty quantification. Instead of relying on traditional microscopes with lenses and complex optical alignments, the UCLA team built a compact system that captures diffraction patterns of immunohistochemically stained tissue samples, then uses AI to reconstruct and analyze what it sees. The key innovation is that the AI doesn't just make a diagnosis — it also tells doctors how confident it is in that diagnosis. When confidence is low, the system flags the case for a pathologist to review by hand.
The team tested their platform on 412 breast tissue samples, focusing on HER2 scoring, a critical biomarker that directly shapes treatment decisions for breast cancer patients. The results were striking: 84.9% accuracy for the full four-class HER2 scoring system, and 94.8% accuracy for the clinically relevant binary classification — scores that matched what conventional brightfield microscopy systems achieved, but with cheaper, simpler hardware. More importantly, the uncertainty-guided strategy identified less certain predictions and flagged them for review, achieving a 30.4% error-correction rate overall.
"Reliable uncertainty estimation is a critical component for the safe deployment of AI in health care," Ozcan said. That philosophy runs through the entire platform design. By combining computational imaging, deep learning and uncertainty quantification, the researchers created a system that doesn't just diagnose — it knows when it might be wrong.
The implications extend far beyond HER2 assessment. The UCLA framework provides a scalable imaging-AI paradigm that could evaluate a wide range of biomarkers and digital pathology tasks. Perhaps most significantly, the compact lensfree design means the technology doesn't require the high-end infrastructure that keeps advanced diagnostics out of reach in resource-limited settings. Ozcan's team sees this as foundational to building trustworthy AI-assisted diagnostics for cancer detection, diagnosis and treatment guidance — systems that can work not just in leading academic medical centers, but in clinics and hospitals everywhere.
For breast cancer patients worldwide, the stakes of getting HER2 scoring right are impossibly high. This technology, published in BME Frontiers, moves the needle toward a future where that critical decision rests on AI systems that are both accurate and honest about what they don't know.