An MR breast auto prescription on-site learning system to adapt different scan plan preferences: a proof-of-concept study

Motivation: Hospitals have various scan prescription preferences for the same scan target. Data collection and annotation are difficult for in-house development of auto prescription system, especially for breast MRI scan. Goal(s): Present an on-site learning system that supports auto prescription model learning for MR breast scan preferences. Approach: Simulate on-site learning system in-house. Automatically generate training data by an anatomy detection model to train auto prescription models learning two orientation preferences. Results: The average success rate achieved 95% and maximal angle error in 75% cases was less than intra-operator error when training data reached 60 patients. Impact: The proof-of-concept study verifies the proposed on-site learning system's feasibility by in-house simulation. The system could automatically train auto prescription models adapting to different scan orientation preferences, and it is expected to be extended to other scan ROI parameters.