Deep Learning--Based Automatic Frequency Offset Selection for Frequency Scout in Cardiac Cine MRI at 3.0T

Need statement: In Cardiovascular magnetic resonance (CMR) at 3.0 Tesla, bSSFP cine imaging is important for
evaluating myocardial function and structure. These sequences provide high contrast and
temporal resolution.However, increased B0 inhomogeneity and frequency sensitivity at 3.0T
can lead to prominent off-resonance effects, resulting in banding artifacts.These artifacts
can affect diagnostic details, particularly in regions with different magnetic suspectibility
e.g. the region near myoacardal tissue interfaces or metallic implants. Current practices
rely on manual selection of the optimal centre frequency offset from scout images acquired across
a range of frequency offsets. This process is subjective and time-consuming, and likely to vary
between operators. This process is subjective and time-consuming, and likely to vary between
operators. This inefficiency affects patients requiring 3.0T imaging, and those with implanted devices where the
severity of the artifact is intensified, and clinical settings where reliable imaging
is essential for efficient throughput.This lack of dynamic method to
set frequency gives rise to additional time in scanning, causing discomfort to patients and
sometimes errors in diagnosis, thus emphasizing a need for deep learning based approach to
optimize frequency offset selection in 3.0T cine imaging.

Measurable Goal: Our goal would be to achieve an accuracy greater than 85%. Our method would work on multiple

planes including long axis(4ch, 3ch, 2 ch), Short axis and other planes.

Current vs. Goal Capabilities: Currently manual practices for selecting frequency offset include visual inspection of images

obtained from frequency scout scan.The operator then manually uses the frequency offset

with the best signal and least artifact based on his own judgement for CINE imaging.This

process is challenged by subjectivity, time consumption, and sensitivity to field inhomo-

geneities. Our approach would use deep learning to automate this process. Our final goal

would be to identify the optimal offset with minimal artifacts, guided by expert-labeled

ground truths or automated quality metrics (e.g., SNR, sharpness).

Must-Have Features: Our approach would include a pipeline having automatic ROI(Region of Interest) segmen-

tation to accurately localize the left ventricle in each mri slice using a segmentation model,

followed by preprocessing and ROI extraction to enhance generalization. The final step,

frequency tuning, analyzes a stack of 7-16 ROI images across different frequency offsets to

identify the optimal offset with minimal artifacts.

Nice-to-Have Features: Our approach could include the capability to segment all the images simultaneously, enhanc-

ing efficiency across multiple slices. The method could handle respiratory variations, improv-

ing reliability in free-breathing exams. It also might extend to support explainable Artificial Intelligence (XAI)

to provide insights into decision-making processes, aiding clinician diagnostic value and interpretation.