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

Area of Science:

• Biomedical Optics
• Medical Imaging
• Machine Learning

Background:

• Polarization-sensitive optical coherence tomography (PS-OCT) offers rich data for tissue characterization.
• Integrating multi-contrast PS-OCT information may enhance diagnostic accuracy for epidural tissues.
• Deep learning approaches are increasingly used for complex medical image analysis.

Purpose of the Study:

• To evaluate deep learning fusion strategies for multi-contrast PS-OCT data in epidural tissue classification.
• To compare the performance of six different fusion methods against single-channel PS-OCT baselines.
• To assess the impact of fusion on classification accuracy and inter-subject variability.

Main Methods:

• Utilized four PS-OCT channels: intensity, phase retardation, degree of polarization uniformity (DOPU), and optic axis.
• Applied six deep learning fusion strategies to porcine and human spinal specimens.
• Benchmarked fusion methods against single-channel PS-OCT using nested cross-validation and cross-testing.

Main Results:

• Multi-channel fusion consistently improved accuracy and reduced subject-to-subject variability compared to single-channel methods.
• Probability averaging fusion achieved the highest mean validation accuracy on human data (98.26%) and porcine data (93.07%).
• Fusion methods significantly reduced between-subject variance in human data (Levene's p = 0.0108) without prior channel selection.

Conclusions:

• Deep learning fusion of multi-contrast PS-OCT data enhances epidural tissue classification robustness and accuracy.
• Probability averaging presents an effective and computationally efficient fusion strategy, requiring no additional training.
• Multi-channel fusion offers a more reliable approach for epidural tissue classification, reducing diagnostic variability.