Boundary Aware Semantic Segmentation using Pyramid-dilated Dense U-Net for Lung Segmentation in Computed Tomography...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate lung segmentation in CT images is crucial for diagnosing various respiratory conditions.
• Existing methods struggle with pathological lung images containing nodules, cavities, or consolidation.
• Deep learning offers potential for robust and automated segmentation solutions.

Purpose of the Study:

• To propose an efficient and accurate deep learning model for robust lung segmentation from CT images.
• To address the challenge of segmenting lungs with abnormalities like nodules, cavities, and consolidation.
• To evaluate the effectiveness of a novel segmentation architecture and loss function.

Main Methods:

• Development of a pyramid-dilated dense U-Net (PDD-U-Net) model incorporating pyramid-dilated convolution blocks.
• Integration of shallow and deep stream features within a nested U-Net decoder for enhanced segmentation.
• Investigation of three loss functions, with a focus on a shape-aware loss function for precise boundary delineation.
• Testing the model on the Lung CT Segmentation Challenge (LCTSC) and LIDC-IDRI datasets.

Main Results:

• The PDD-U-Net model achieved high segmentation accuracy on both standard and pathological lung CT images.
• Achieved a Dice coefficient of 0.983 on the LIDC-IDRI dataset and 0.994 on the LCTSC dataset.
• Demonstrated superior performance compared to other segmentation methods evaluated.

Conclusions:

• The PDD-U-Net model with shape-aware loss is an effective and accurate method for lung segmentation in CT images, handling abnormalities well.
• The model's architecture, combining pyramid-dilated convolutions and a nested U-Net decoder, enhances segmentation precision.
• This approach holds significant potential for improving computer-aided diagnosis systems through rapid and accurate lung region analysis.