Assistance technique for Fiducial Localization in MRI images with Subslice Classification Neural Network

Image-guided neurosurgeries(IGNS) require patient-to-image registration for treatment planning. As the patient head-surface has very few recognizable features, adhesive markers called Fiducials are affixed on the patient head and are localized to establish the registration. While various approaches have been proposed in literature to automate the localization process, this research work has faced roadblocks while constructing a proof- of-concept over these approaches: These techniques assume access to a ground-truth annotated dataset which either requires availability of the same in the public domain or via an IGNS expert. To the best of our knowledge, there is no publicly available annotated dataset for evaluation of fiducial localization algorithms. Similarly, access to IGNS experts is also a strong assumption which greatly deters the entry of researchers willing to address the challenge of fiducial localization. The proposed study has attempted to address these challenges by reporting a simple ground-truth annotation procedure with which the study have applied over 50 preoperative MRI volumes, for annotation of fiducial centers. Further, this research work proposes an approach based on a sub-slice classification neural network to extract region-of-interest in a MRI slice over which clustering and edge-detection algorithms are applied to localize the fiducials. Our approach achieves mean 4mm Fiducial localization error(FLE) and around 70% precision and recall in predicting fiducial center coordinates, thereby achieving significant improvement over our earlier approach based on classification of detected corners in MRI slices. Also, this research study has observed a 30% reduction in the time required for fiducial localization in a randomly selected volume using the proposed approach with respect to manual localization. While the clinical requirements for fiducial localization error are in the order of 1mm, our technique can serve as a reasonable initial guess, thereby assisting neurosurgeons in the pre-operative workflow.