Hybrid Deep Learning and Model-Based Needle Shape Prediction.

Area of Science:

• Medical Robotics
• Surgical Navigation
• Computational Mechanics

Background:

• Minimally-invasive surgery for prostate cancer often uses flexible bevel tip needles for precise steering.
• Accurate intra-operative prediction of needle trajectory is essential to avoid sensitive structures and reduce re-insertions.
• Predicting needle path during insertion is challenging due to unpredictable needle-tissue interactions.

Purpose of the Study:

• To develop and validate a novel hybrid approach for intra-operative needle shape prediction.
• To leverage a Lie-group theoretic model for accurate needle shape representation.
• To introduce a self-supervised learning method for training prediction networks without prior data.

Main Methods:

• A hybrid deep learning and model-based approach was developed, integrating a validated Lie-group model for needle shape.
• A novel self-supervised learning method was employed for network training, enabling data-scarce scenarios and transfer learning.
• Needle shape prediction was tested in homogeneous phantom tissues for C- and S-shaped insertions.

Main Results:

• The hybrid method achieved an average root-mean-square prediction error of 1.03 mm.
• The system was validated on a dataset of approximately 3,000 prediction samples.
• Predictions were made for needle insertions up to 110 mm in length.

Conclusions:

• The presented hybrid approach offers a significant advancement in intra-operative needle shape prediction.
• The self-supervised learning method facilitates robust network training, even with limited data.
• This technology has the potential to enhance the accuracy and efficiency of minimally-invasive prostate cancer interventions.