Low dose threshold for measuring cardiac functional metrics using four-dimensional CT with deep learning.

Area of Science:

• Medical imaging
• Artificial intelligence
• Cardiovascular imaging

Background:

• Four-dimensional CT (4DCT) is vital for cardiac imaging and prognosis but poses radiation dose concerns.
• Reducing radiation exposure during 4DCT is critical for patient safety.
• Deep learning (DL) offers a potential solution for dose reduction in cardiac CT.

Purpose of the Study:

• To investigate deep learning-based segmentation for dose reduction in 4DCT cardiac imaging.
• To evaluate the impact of dose reduction on functional cardiac metrics.
• To assess the performance of DL models trained with standard-dose and low-dose data.

Main Methods:

• Developed a 3D residual U-Net for segmenting left ventricle (LV) myocardium and blood pool.
• Trained two DL networks: Standard DL (SD data only) and Noise-Robust DL (SD + low-dose data).
• Measured functional cardiac metrics (EF, GLS, CS, WT) on 250 4DCT volumes at varying dose levels.

Main Results:

• DL-derived functional metrics closely matched expert manual analysis for standard-dose images.
• Standard-DL showed minimal differences (<0.8%) in EF, GLS, CS with ~76% dose reduction.
• Noise-Robust DL maintained acceptable performance even at significantly reduced doses (50 mA).

Conclusions:

• Radiation dose in 4DCT can be reduced by an average factor of 5 with minimal impact on global functional metrics.
• DL segmentation techniques enable significant dose reduction while preserving diagnostic accuracy.
• Training DL models with emulated low-dose data enhances robustness to reduced image quality.