Rapid AI OCT System Detects Malignant Liver Tissue Intraoperatively

Liver cancer commonly requires resection, yet confirming complete tumor removal relies on frozen-section analysis that prolongs anesthesia and ties up operating rooms. Longer cases increase risks for bleeding, infection, and airway or hemodynamic instability, and they add burdens for surgical teams. Faster, dependable intraoperative assessment would improve safety and efficiency. To help address this challenge, researchers have now combined optical coherence tomography with artificial intelligence to accelerate analysis of liver specimens.

The approach pairs optical coherence tomography (OCT) with an anomaly detection algorithm developed by teams at University Hospital RWTH Aachen, the Fraunhofer Institute for Production Technology (IPT), and Fraunhofer Austria Research. OCT provides volumetric, cross‑sectional views of tissue and is already established in ophthalmology. The project adapts this imaging modality for intraoperative use in liver surgery and applies machine learning to distinguish normal parenchyma from malignant tissue.

OCT generates three-dimensional scans using light waves, producing detailed images within seconds. The anomaly detection model is trained only on scans of normal liver parenchyma, enabling it to flag departures from the learned distribution as suspicious. This training paradigm suits settings where nonmalignant samples outnumber malignant ones and allows markedly faster model development than traditional supervised classifiers.

Under laboratory conditions, the collaborators acquired 173 OCT scans from 69 patients at University Hospital RWTH Aachen, including 88 scans of normal liver parenchyma and 85 scans of various tumor types. Applying anomaly detection to human liver OCT images represents a first-of-its-kind combination in this context. The work demonstrates feasibility for rapid, intraoperative decision support.

Classification is produced within seconds, indicating whether a scan reflects normal or tumor tissue, with subsequent confirmation by standard histopathology. Using the available data, the model achieved a mean accuracy of 81%. Depending on tumor type, observed accuracies reached 94.3%, 84.5%, and 65.9%. The findings were published in Scientific Reports.

Next steps include translating the method from laboratory to operating room conditions, miniaturizing the sensor, and integrating the system into surgical workflows as a complement to histopathology. These developments aim to support quicker resections and reduce strain on perioperative teams without sacrificing diagnostic assurance.

“What is special about this method is that the model is trained exclusively on good examples—that is, scans of normal liver parenchyma. The method then reliably detects deviations from this distribution. Using the available data, we achieved a mean accuracy of 81% and have thereby demonstrated that anomaly detection is well-suited as a decision-support tool in this context. Our work has provided the proof of concept,” said Ulrich Krispel, an anomaly detection expert at Fraunhofer Austria.

Related Links
Fraunhofer Austria