Contour-aware multi-label chest X-ray organ segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of chest organs in X-ray images is crucial for diagnosing lung diseases and assessing organ size.
• Deep convolutional neural networks (CNNs) are state-of-the-art for image segmentation tasks.

Purpose of the Study:

• To investigate the benefits of augmenting CNNs with organ contour information for chest X-ray segmentation.
• To evaluate the performance of contour-augmented CNNs on segmenting lung fields, heart, and clavicles.

Main Methods:

• Three CNN architectures (UNet, LinkNet with ResNeXt, Tiramisu with DenseNet) were augmented with organ contour data.
• Architectures were trained on both ground-truth segmentation masks and corresponding organ contours.
• Performance was compared against contour-free versions and 20 existing lung field segmentation algorithms.

Main Results:

• Contour-aware segmentation significantly improved performance across all evaluated CNN architectures.
• The UNet architecture with ResNeXt50 encoder and contour-aware approach achieved the highest segmentation performance.
• Achieved Jaccard overlap coefficients of 0.971 (lung fields), 0.933 (heart), and 0.903 (clavicles).

Conclusions:

• Augmenting CNNs with organ contour information enhances segmentation accuracy in chest X-rays.
• The proposed contour-aware approach outperformed all existing methods on a public chest X-ray dataset.