Deep learning fusion of multi-channel imaging from polarization-sensitive optical coherence tomography

Multi-contrast polarization-sensitive optical coherence tomography (PS-OCT) provides complementary structural and polarization information that may improve epidural tissue classification. Here, we evaluated deep learning fusion of four PS-OCT channels, including intensity, phase retardation, degree of polarization uniformity (DOPU), and optic axis, using porcine (n = 6) and human (n = 5) spinal specimens. We benchmarked six multi-channel fusion strategies: Probability averaging, feature concatenation, trainable weighted output, shared-stage resnet, merged multi-channel input, and pooled data. Across subject-level nested cross-validation, multi-channel methods achieved modest but consistent accuracy improvements over the best single-channel baselines while reducing subject-to-subject variability. On porcine data, Probability Averaging increased mean validation accuracy by 3.46% (93.07% vs. 89.61% for the best single-channel baseline). On human data, fusion methods maintained the high single-channel baseline (approximately 97% to 98%) while modestly improving stability, with probability averaging achieving the highest mean validation accuracy (98.26%). In cross-testing, trainable weighted output achieved 92.32% versus 91.53% for the best porcine single-channel baseline, and probability averaging achieved 97.87% versus 97.75% for the best human single-channel baseline. On the human cohort, the gain in mean accuracy was small, but multi-channel fusion produced a statistically significant reduction in between-subject variance (Levene's p = 0.0108) and removed the need to know in advance which single channel would generalize best. Overall, multi-channel fusion improved classification performance and robustness, with probability averaging offering a favorable balance between accuracy and complexity because it requires no additional training beyond single-channel models.