Towards rapid differentiation of liver malignancies using optical coherence tomography and deep learning

Background: Liver cancer, in particular hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (iCCA), and colorectal liver metastases (CRLM), pose significant global health challenges due to high mortality rates. Complete surgical resection leads to the best oncological outcomes, but intraoperative identification of negative resection margins is a time-consuming process, relying on frozen section analysis. Optical Coherence Tomography (OCT) combined with deep learning could provide a real-time, efficient method for resection margin examination.
Methods: This study utilized OCT imaging to scan fresh liver resection specimens ex-vivo. Theimages were processed using convolutional neural networks (CNNs) trained to differentiate between HCC, iCCA, CRLM, and liver parenchyma. ResNet50 and Xception architectures were applied and compared using five-fold cross-validation. To evaluate the performance of the binary classifications, we used the area under the receiver operating characteristic (AUROC).
Results: A total of 205 three-dimensional C-Scans (419,840 2D images) collected from 91 patients were included. Of these, 17 were iCCA, 24 HCC, 59 CRLM, and 105 non-tumor liver parenchyma. Our method achieved high performance in tissue classification, showing an average AUROC of 0.9041 for distinguishing liver parenchyma from malignancies. Differentiating between malignancies proved to be more challenging. Binary classifications trained on pairs of tissue types exhibited varying performances for the different tissue types.
Conclusion: The integration of OCT with deep learning demonstrates a promising approach for the rapid detection of liver malignancies during surgery. This method has the potential to decrease surgery duration, thereby reducing associated risks and improving the experiences of both patients and surgeons.