Deep-learning-aided sarcoma biopsy guidance using forward-view endoscopic optical coherence tomography

Sarcoma diagnosis remains challenging due to tumor heterogeneity and the risk of sampling errors during biopsy, as tissue misidentification can result in non-diagnostic outcomes and repeated procedures. This study specifically targets undifferentiated pleomorphic sarcoma (UPS) and myxofibrosarcoma, imaged in specimens from the human arms and legs. Optical coherence tomography (OCT) offers high-resolution, real-time microstructural imaging and has shown strong potential for image-guided oncology. In this study, we benchmarked five 2D deep learning models and one custom 3D convolutional neural network (CNN) for normal versus tumor tissues classification using endoscopic OCT. The 2D models were applied to cross-sectional 2D images while the 3D CNN operated on volumetric images, with both evaluated across the Intensity and polarization-sensitive OCT channels (PS-OCT). 3D CNN reached the highest accuracy of 92.67% on Intensity, while PS-OCT channels benefited more from modern 2D architectures and 3D volumetric modeling, with performance gains up to 9.63% for the optic axis. These results demonstrated the feasibility of endoscopic OCT as a real-time, label-free tissue characterization tool during biopsy of these sarcoma tissues. The integration of deep learning with forward-viewing endoscopic OCT offers a promising pathway toward more accurate needle guidance and reduced diagnostic failure rates and established a reproducible benchmarking framework for future OCT-based detection and analysis of sarcoma tissues.