Uncertainty-Aware AI Platform Supports Automated HER2 Assessment in Breast Cancer

While artificial intelligence (AI)-enabled digital pathology offers scalable, quantitative analysis, adoption remains limited by costly imaging systems and the need for reliable measures of prediction confidence. New findings demonstrate a compact imaging and AI platform that automates HER2 assessment while flagging cases that may need additional review.

University of California, Los Angeles (UCLA) researchers developed an uncertainty-aware computational pathology platform that combines lensfree holographic imaging with deep learning to automate HER2 scoring in breast cancer tissue. The compact lensfree system captures diffraction patterns over a large field of view from immunohistochemically stained samples, then computationally reconstructs the specimen for AI-based analysis. To bolster reliability, the workflow integrates a confidence-aware deep learning framework with Bayesian uncertainty quantification that estimates prediction confidence and flags low-certainty cases for additional review.

In a blinded test set of 412 breast tissue samples, the platform achieved 84.9% accuracy for four-class HER2 scoring and 94.8% accuracy for clinically relevant binary classification. Selectively identifying less-certain predictions yielded an overall error-correction rate of 30.4%, providing an added layer of safety for clinical decision support by mitigating low-confidence outputs. Despite simplified optical hardware, performance was comparable to conventional brightfield microscopy–based digital pathology while enabling high-throughput imaging across a large sample area.

The study is published in BME Frontiers on May 21, 2026. The researchers note that this lensfree imaging–AI paradigm can extend beyond HER2 to other biomarker evaluation tasks and digital pathology applications. By pairing uncertainty-aware AI with compact computational imaging hardware, the approach is presented as a route toward more accessible and trustworthy AI-assisted diagnostics for cancer detection, diagnosis and treatment guidance.

"Reliable uncertainty estimation is a critical component for the safe deployment of AI in health care," said Aydogan Ozcan, Chancellor’s Professor at UCLA and lead author of the study. "By combining computational imaging, deep learning and uncertainty quantification, we aim to create accessible and trustworthy diagnostic technologies that can support pathology workflows in both advanced and resource-limited health care settings." 

Related Links
University of California, Los Angeles