Fully-automated deep learning-based flow quantification of 2D CINE phase contrast MRI

Area of Science:

• Cardiovascular Imaging
• Medical Artificial Intelligence
• Radiology

Background:

• Time-resolved, 2D-phase-contrast MRI (2D-CINE-PC-MRI) is crucial for in vivo blood flow analysis.
• Accurate vessel contour delineation (VCD) is essential for reliable 2D-CINE-PC-MRI results.
• Manual analysis (MA) and semi-automated methods can be time-consuming and prone to variability.

Purpose of the Study:

• To evaluate a fully-automated deep learning (DL) application for VCD and blood flow quantification.
• To compare the performance of DL analysis against manual analysis (MA) and corrected semi-automated analysis (corSAA).
• To assess the accuracy and efficiency of DL in analyzing 2D-CINE-PC-MRI data.

Main Methods:

• 97 patients with 2D-CINE-PC-MRI at the sinotubular junction (STJ) and 28 at the main pulmonary artery (PA) were included.
• A cardiovascular radiologist performed MA (reference) and corSAA; DL performed automated VCD and flow quantification (net flow [NF] and peak velocity [PV]).
• Contour accuracy was assessed using Dice similarity coefficients (DSC); discrepant cases were reviewed.

Main Results:

• DL was successfully applied to 97% of imaging series, demonstrating good to excellent performance (mean DSC: 0.91 at STJ, 0.85 at PA).
• Flow quantification showed similar net flow between DL and human assessments (p > 0.05).
• DL analysis was accurate in 93.4% of cases, with instantaneous results compared to manual assessments.

Conclusions:

• Fully-automated DL analysis of 2D-CINE-PC-MRI provides expert-level flow quantification at STJ and PA in over 93% of cases.
• DL offers instantaneous results, significantly improving efficiency over manual methods.
• The evaluated DL tool demonstrates usability and potential for integration into daily clinical practice.