Real-time AI image reconstruction and instrument tracking in MR-guided biopsies

Motivation: Interventional MRI struggles with speed and efficiency. Goal(s): To accelerate transrectal in-bore MR-guided biopsy procedures through undersampled image reconstruction and instrument tracking using a novel deep learning AI approach. Approach: Image reconstruction and instrument segmentation models were trained using 8457 synthetically undersampled clinical MR-guided biopsy scans from 1289 patients. The models were tested on 5 prospectively undersampled MR-guided biopsy dynamic k-space datasets, evaluating the needle guide tip prediction error and the failure rate of needle guide prediction at increasing levels of undersampling. Results: We found stable performance with up to 16x undersampling. Impact: A deep-learning temporal model utilizing spatiotemporal information achieved up to 16x undersampling rates for MR-guided prostate biopsy scans while maintaining accurate instrument tip positioning. This could enable real-time instrument tracking in interventional tasks, improving efficiency and usability.