Enabling a one touch MR patient setup using RIS Interpretation and 3D Camera

Motivation: Making the MRI patient setup as simple as laying the patient on the table. Goal(s): AI based solution integrates RIS (Radiology Information System) with in room camera based key-points & coil placement on the patient to set the landmark automatically. Approach: Solution which incorporates TinyBERT based inference to get user intent from study descriptions (RIS), EfficientDet based generic coil detection and Detectron2 based human keypoint detection on images captured through 3D camera Results: Trained models achieve 92% accuracy for RIS interpretation and ~96% precision and recall for coil detection. Workflow driven by the outputs proven to help coil centering. Impact: The trained Coil detection and RIS interpretation model aids in interpreting the scan intent thereby enabling proper positioning of patient, coil and automatic landmarking, saving time and avoiding repeat scans.