Deep learning can yield clinically useful right ventricular segmentations faster than fully manual analysis.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Manual delineation of right ventricular (RV) volumes from cardiac magnetic resonance (CMR) images is time-consuming.
• Deep learning (DL) methods show potential for automating RV delineations, but clinical practice acceleration is understudied.

Purpose of the Study:

• To develop and validate a clinical pipeline for DL-based RV delineations.
• To assess the time reduction achievable compared to manual methods in clinical practice.

Main Methods:

• A DL model was trained on 1114 subjects' short-axis CMR scans.
• Two observers assessed 50 clinical scans, rating automated delineations (A: clinical use, B: minor correction, C: major correction).
• Time for manual correction (B/C) vs. fully manual delineation was measured.

Main Results:

• 58% of automated delineations were rated sufficient for clinical use (A), 42% required minor correction (B), and none required major correction (C).
• Automated delineation took 2 seconds, manual correction took 2 minutes, and fully manual delineation took 6 minutes.
• An 87% time reduction was achieved with the DL pipeline.

Conclusions:

• The DL-based pipeline substantially reduces time for clinically applicable RV delineations.
• This approach can accelerate right ventricular assessments in clinical practice compared to fully manual analysis.
• Generalizability may be limited to clinics using similar RV delineation guidelines.