Relational Modeling for Robust and Efficient Pulmonary Lobe Segmentation in CT Scans.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pulmonary lobe segmentation in CT scans is vital for regional lung disease assessment.
• Convolutional neural networks show promise but struggle with complex structural relationships.
• Accurate delineation is challenging in diseases like COVID-19 and COPD.

Purpose of the Study:

• To propose a novel relational deep learning approach (RTSU-Net) for improved pulmonary lobe segmentation.
• To leverage structured relationships and non-local neural network modules for enhanced accuracy.
• To evaluate RTSU-Net's performance on diverse lung conditions, including COVID-19.

Main Methods:

• Developed RTSU-Net, incorporating a novel non-local neural network module to learn visual and geometric relationships.
• Trained and validated RTSU-Net on the COPDGene dataset (5000 subjects).
• Applied transfer learning using COPDGene pre-trained models to evaluate RTSU-Net on COVID-19 suspect data (470 subjects).

Main Results:

• RTSU-Net demonstrated superior performance compared to three baseline methods.
• The model showed robust performance even in cases with severe lung infections from COVID-19.
• Learned self-attention weights effectively captured inter-lobe structural dependencies.

Conclusions:

• The proposed RTSU-Net effectively captures critical structural relationships for accurate pulmonary lobe segmentation.
• This relational approach offers significant improvements over traditional convolutional methods.
• RTSU-Net shows promise for clinical applications in diagnosing and monitoring lung diseases like COVID-19 and COPD.