TVnet: Automated Time-Resolved Tracking of the Tricuspid Valve Plane in MRI Long-Axis Cine Images with a Dual-Stage...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Tricuspid valve (TV) tracking in cardiac MRI is crucial for evaluating right ventricular dysfunction, prevalent in congenital heart disease and pulmonary hypertension.
• Manual TV annotation is challenging due to rapid motion and poor visibility, making it time-consuming and difficult.

Purpose of the Study:

• To develop and validate TVnet, a dual-stage deep learning pipeline for automated tricuspid valve tracking and tricuspid annular plane systolic excursion (TAPSE) derivation from cardiac MRI.
• To improve the accuracy and efficiency of TV annotation compared to traditional methods.

Main Methods:

• A dual-stage deep learning pipeline (TVnet) utilizing ResNet-50 and automated image linear transformation was developed.
• Stage 1 performed coarse TV point detection; Stage 2 standardized image orientation and accurately located TV points.
• The model was trained and validated on 4170 MRI images from 140 patients.

Main Results:

• The standardized TVnet model achieved a significantly lower error (2.4 ± 1.7 mm) compared to the baseline (4.0 ± 3.1 mm).
• The standardized model demonstrated excellent agreement with clinical metrics (ICC = 0.94), comparable to inter-observer variability (ICC = 0.92).

Conclusions:

• TVnet substantially enhances tricuspid valve annotation accuracy in cardiac MRI.
• This automated approach facilitates reliable assessment of right ventricular dysfunction, offering a valuable tool for clinical practice.