Related Experiment Video
Updated: Aug 8, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Integrating physics guidance into deep learning for cine cardiovascular segmentation
My Anh Tran-Nguyen1, Quoc Khanh Luong1, Minh Bao Kha2
1Hanoi University of Science and Technology, Hanoi, Viet Nam.
Summary
PGE-UNet enhances cardiovascular magnetic resonance (CMR) image analysis by incorporating physical constraints into deep learning models. This physics-regularized approach improves segmentation accuracy and plausibility for faster, more reliable cardiac function assessment.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Biophysics
Background:
- Cine cardiovascular magnetic resonance (CMR) is crucial for quantifying cardiac function.
- Current deep learning segmentation models lack physical constraints, limiting plausibility with acquisition variability.
- There is a need for efficient and reproducible segmentation methods in routine CMR workflows.
Purpose of the Study:
- To develop an efficient deep segmentation framework (PGE-UNet) for cine CMR.
- To incorporate radiofrequency-related physical constraints for plausible segmentation outputs.
- To enable physics-regularized training without additional CMR acquisition sequences.
Main Methods:
- Proposed PGE-UNet framework with a Noise-Aware Encoder for robust feature encoding.
- Implemented a Physics-Regularized Decoder enforcing Maxwell-Helmholtz consistency.
- Introduced a simulator-based procedure to synthesize pseudo transmit-field priors for training.
Main Results:
- Achieved high accuracy on public cine CMR benchmarks: DICE of 0.9152 (ACDC), 0.8605 (M&Ms), 0.8548 (SCD).
- Demonstrated lightweight model with 1.60 million parameters.
- Recorded fast CPU-only inference latency of 72.38 ms per case.
Conclusions:
- PGE-UNet offers an efficient and accurate solution for cine CMR segmentation.
- The physics-regularized approach enhances prediction plausibility and robustness.
- The model is suitable for time-sensitive and resource-constrained clinical environments.
