A data-driven semantic segmentation model for direct cardiac functional analysis based on undersampled radial MR cine...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Cardiac cine MRI acquisition and manual analysis are time-consuming processes.
• Accelerating MRI acquisition and postprocessing is crucial for clinical efficiency.

Purpose of the Study:

• To train and evaluate a 3D artificial neural network for semantic segmentation of radially undersampled cardiac MRI.
• To accelerate cardiac MRI scan time and postprocessing using deep learning.

Main Methods:

• A 3D U-Net architecture was pretrained on Cartesian cardiac MRI data.
• Transfer learning was applied using non-Cartesian radial cine MRI data for optimization.
• Performance was evaluated using Dice Similarity Coefficient (DSC) at various undersampling levels.

Main Results:

• Without transfer learning, the pretrained model showed moderate performance (max DSC 0.87 for left ventricle).
• After transfer learning, the model achieved human-level performance even with high undersampling rates (e.g., DSC 0.95 for left ventricle at P=33).
• The deep learning approach showed no significant difference compared to segmentations from fully sampled data.

Conclusions:

• A 3D U-Net architecture effectively performs semantic segmentation of radially undersampled cine MRI.
• This deep learning method achieves performance comparable to human experts on fully sampled data.
• The approach jointly accelerates MRI acquisition and manual image analysis.