Spatio-temporal deep learning models for tip force estimation during needle insertion.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Machine Learning

Background:

• Precise needle placement is crucial for clinical procedures like brachytherapy and biopsy.
• Tissue deformation and needle deflection due to forces can lead to misplacement and injury.
• Existing methods for estimating needle tip forces face challenges in sensor integration and calibration.

Purpose of the Study:

• To develop and validate a novel fiber-optic needle tip force sensor.
• To investigate deep learning approaches for calibrating the sensor.
• To propose a new convGRU-CNN architecture for processing spatio-temporal image data.

Main Methods:

• A fiber-optic sensor design utilizing a single OCT fiber to image epoxy layer deformation below the needle tip.
• Acquisition of 1D depth profiles representing tissue deformation.
• Application of deep learning, including a novel convGRU-CNN architecture, for force estimation and calibration.

Main Results:

• The sensor's operating range is adaptable by altering epoxy layer stiffness.
• The convGRU-CNN architecture achieved a mean absolute error and a cross-correlation coefficient of 0.9997, outperforming other methods.
• Successful ex vivo application demonstrated in human prostate tissue.

Conclusions:

• The OCT-based fiber-optic sensor offers a viable solution for needle tip force estimation.
• Deep learning models effectively utilize spatio-temporal image data for force measurement.
• The convGRU-CNN architecture shows promise for similar spatio-temporal learning tasks.