RPLS-Net: pulmonary lobe segmentation based on 3D fully convolutional networks and multi-task learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate pulmonary lobe segmentation is crucial for surgical planning and disease analysis in Computer Aided Diagnosis systems.
• Challenges in segmentation include unclear lobe borders due to anatomical variations and disease-related obstructions.

Purpose of the Study:

• To propose a Regularized Pulmonary Lobe Segmentation Network for precise segmentation of lung lobes and their borders.
• To address the difficulties in segmenting unclear pulmonary lobe boundaries.

Main Methods:

• Utilized a 3D fully convolutional network for feature extraction from CT images.
• Employed multi-task learning for joint segmentation of lobes and borders.
• Introduced a 3D depth-wise separable de-convolution block for efficient deep supervision and a hybrid loss function.

Main Results:

• Achieved a mean Dice coefficient of 0.9421 and an average symmetric surface distance of 1.3546 mm.
• Demonstrated comparable performance to state-of-the-art methods.
• Successfully segmented voxels near lung walls and fissures.

Conclusions:

• A 3D fully convolutional network framework effectively segments pulmonary lobes and borders in chest CT images.
• The proposed approach shows significant effectiveness in segmenting both lung tissues and their boundaries.