Robust 3D Landmark Detection Framework for One-Stop Automated Pelvic MRI Prescription

Motivation: Automating the prescription of uterine body and cervix in the short-axis to assess endometrial cancer invasion has been challenging due to anatomical variations from diseases and the necessity of a two-step prescription process involving pre-scan/main-scan. Goal(s): Our goal was to streamline this process into a one-stop automated workflow. Approach: We developed a novel deep learning-based positioning method that detects 3D landmarks from 3D scout pre-scan and T2-weighted sagittal main-scan images. Results: Our method achieved technician acceptance rates of 85.3% for uterine body and 97.1% for cervix prescriptions on a dataset primarily consisting of cases with lesions, demonstrating robustness against various diseases and artifacts. Impact: The automated one-stop workflow enables single-button operation for pelvic MRI, including the challenging short-axis positioning of the uterine body and cervix. It reduces prescription variability among technicians and ensures reproducible imaging, even in anatomically complex cases due to diseases.