gms | German Medical Science

Introduction: Intraoperative differentiation of liver malignancies, including hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (iCCA), and colorectal liver metastases (CRLM) from liver parenchyma is essential to ensure negative resection margins. Current methods, such as frozen section analysis, are time-intensive and can prolong surgery. This study explores the use of optical coherence tomography (OCT) combined with deep learning as a rapid, automated alternative, in an ex vivo setting.

Method: Freshly resected liver tissues (parenchyma, HCC, iCCA, CRLM) from 91 patients were scanned between June 2020 and April 2021, using a table-top spectral-domain OCT device at 1310 nm wavelength. Areas of interest were imaged before formalin fixation and marked for histological examination, providing matched diagnoses for each scan. A stratified 5x5 cross-validation process was used to alternately split the dataset into training, validation and test sets and two convolutional neural networks (CNN - ResNet50, Xception) were trained and validated in binary (parenchyma vs. malignancies, pairwise tumor comparisons) and multi-class classification tasks.

Results: During the study period, 205 three-dimensional scans (of which 17 iCCA, 24 HCC, 59 CRLM, and 105 liver parenchyma) and over 400,000 two-dimensional B-Scans were generated. Differentiating healthy liver parenchyma from malignancies achieved a high average AUROC of 0.91 (Xception). Binary classifications of parenchyma vs iCCA and parenchyma vs CRLM performed particularly well, with AUROCs of 0.90 and 0.96, respectively. Differentiating parenchyma from HCC was more challenging, yielding an AUROC of 0.7. Pairwise tumor-tumor classifications, such as iCCA vs CRLM and iCCA vs. HCC, showed lower AUROCs (0.68–0.71), likely reflecting overlapping morphological features and variability within tumor subtypes. The two CNN performed similarly, with Xception showing slightly better predictive capability overall.

Conclusion: OCT imaging combined with deep learning provides a promising approach for rapid and automated intraoperative differentiation of liver tissue, potentially reducing the reliance on frozen section analysis and shortening surgery duration. This method has the potential to improve patient safety and surgeon efficiency. Further studies are needed to validate these results in vivo and enhance classification accuracy for diverse tumor subtypes.