Automated Placement of Scan and Pre-Scan Volumes for Breast MRI Using a Convolutional Neural Network

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Magnetic Resonance Imaging (MRI)

Background:

• Manual placement of patient-specific imaging and pre-scan volumes in MRI is time-consuming and prone to errors.
• This variability is a bottleneck, especially with the increasing use of abbreviated breast MRI for screening.
• Optimizing image quality and workflow efficiency is crucial in modern breast MRI examinations.

Purpose of the Study:

• To develop and validate an automated approach for placing scan and pre-scan volumes in breast MRI using deep learning.
• To address the limitations of manual volume placement, including time consumption and operator variability.
• To improve the efficiency and consistency of breast MRI procedures, particularly for abbreviated exams.

Main Methods:

• A deep convolutional neural network was trained on 3-plane scout images from 333 clinical breast MRI exams.
• The network predicted both scan volumes and bilateral pre-scan volumes.
• Performance was evaluated using intersection over union (IoU), center distance, and volume size difference compared to manual placements.

Main Results:

• The automated scan volume placement achieved a median 3D IoU of 0.69, with a median location error of 2.7 cm and size error of 2%.
• The pre-scan volume placement yielded a median 3D IoU of 0.68, with a median location error of 1.3 cm and size error of -2%.
• Estimated uncertainties in positioning and volume size ranged from 0.2 to 3.4 cm for both models.

Conclusions:

• An automated, neural network-based approach for scan and pre-scan volume placement in breast MRI is feasible.
• This method demonstrates potential to significantly reduce manual labor and improve consistency in breast MRI protocols.
• Automation of volume placement can streamline abbreviated breast MRI workflows, enhancing screening efficiency and diagnostic accuracy.