Automatic lung segmentation in dynamic thoracic MRI using two-stage deep convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Lung segmentation in dynamic thoracic MRI (dMRI) is crucial for analyzing thoracic structure and function in respiratory disorders.
• Existing segmentation methods, often developed for CT, lack efficiency and robustness for dMRI.
• The unsuitability of traditional methods for large dMRI datasets necessitates novel approaches.

Purpose of the Study:

• To develop a novel automatic lung segmentation approach specifically for dMRI datasets.
• To improve the efficiency and accuracy of lung segmentation in dynamic thoracic imaging.
• To provide a robust solution for analyzing large dMRI datasets in patients with respiratory disorders.

Main Methods:

• A two-stage convolutional neural network (CNN) approach was developed for automatic lung segmentation.
• Pre-processing involved modified min-max normalization to enhance lung-tissue contrast.
• A region of interest (ROI) detection strategy using corner points and CNNs was employed to isolate lung ROIs from sagittal dMRI slices.
• A modified 2D U-Net architecture processed adjacent ROIs for final lung segmentation.

Main Results:

• The proposed method demonstrated high accuracy in lung segmentation for dMRI.
• The approach achieved significant stability in segmenting lung tissue across dMRI datasets.
• Qualitative and quantitative evaluations confirmed the effectiveness of the novel segmentation technique.

Conclusions:

• The presented two-stage CNN approach offers an effective and robust solution for automatic lung segmentation in dMRI.
• This method addresses the limitations of traditional techniques, enabling efficient analysis of large dMRI datasets.
• The findings support the application of this approach for quantitative analysis in patients with respiratory disorders.