Real-Time Deep-Learning Image Reconstruction and Instrument Tracking in MR-Guided Biopsies.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Biopsy Techniques

Background:

• Transrectal in-bore MR-guided biopsy (MRGB) offers accuracy but is limited by long scan times.
• Faster imaging is crucial for improving MRGB workflow and enabling real-time instrument tracking.
• Current acceleration methods often rely on simulated data and lack clinical validation.

Purpose of the Study:

• To accelerate MRGB procedures using deep learning for image reconstruction and instrument tracking.
• To train models on multi-slice MR DICOM images and evaluate them on raw k-space acquisitions.
• To enable faster and more efficient MR-guided biopsies.

Main Methods:

• A prospective feasibility study involving 1289 patients for model training and 8 patients for testing.
• Deep learning models were trained on 8464 MRGB scans for segmentation and reconstruction.
• Instrument tip prediction (ITP) error was used to assess tracking accuracy, with feasibility measured by the proportion of frames with <5mm error.

Main Results:

• The reconstruction model achieved a mean ITP error of 1.55 ± 1.01 mm on fully sampled scans.
• High ITP success rates of 97.5% at 8x and 92.5% at 16x undersampling were maintained.
• Performance showed a decline at 18x undersampling, indicating the limits of the current model.

Conclusions:

• Deep learning models demonstrate stable needle guide tip prediction accuracy in MRGB.
• The reconstruction model is robust for tracking at high undersampling rates.
• This AI-driven approach shows promise for accelerating MRGB procedures.