Tissue Classification During Needle Insertion Using Self-Supervised Contrastive Learning and Optical Coherence...

Area of Science:

• Medical Imaging
• Machine Learning in Medicine
• Biomedical Engineering

Background:

• Accurate needle positioning is critical in medical procedures like epidural anesthesia.
• Physicians currently rely on tactile feedback and anatomical knowledge for needle guidance.
• Objective tissue identification at the needle tip can provide crucial supplementary feedback.

Purpose of the Study:

• To develop a deep neural network (DNN) for classifying tissue types encountered during needle insertion.
• To evaluate the DNN's performance using phase and intensity data from optical coherence tomography (OCT) signals.
• To introduce and assess a novel contrastive pretraining strategy for improving DNN performance in limited labeled dataset scenarios.

Main Methods:

• Acquisition of complex OCT signals (phase and intensity data) at the needle tip during insertion.
• Development and application of a deep neural network for tissue classification based on OCT data.
• Implementation of a novel contrastive pretraining strategy to learn invariant representations from phase and intensity data.

Main Results:

• The proposed contrastive pretraining strategy significantly improved the DNN's F1 score to 0.84±0.10 using only 10% of the training data.
• The model without pretraining achieved a lower F1 score of 0.60±0.07 under similar limited data conditions.
• Analysis revealed the individual contributions of phase and intensity data to tissue classification accuracy.

Conclusions:

• Deep neural networks, enhanced by contrastive pretraining, show significant promise for real-time tissue classification during needle insertion.
• The proposed pretraining method is effective in improving DNN performance, particularly in low-data regimes.
• This approach offers a potential pathway to enhance needle guidance and improve patient safety in various medical interventions.