Dilated conditional GAN for bone suppression in chest radiographs with enforced semantic features

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Bone structures in chest X-rays can obscure lung pathologies.
• Accurate computer-aided diagnosis requires clear visualization of lung tissue.

Purpose of the Study:

• To develop an algorithm for suppressing bone structures in X-ray images.
• To improve the clarity of lung tissue for better disease detection, such as lung nodules.

Main Methods:

• A conditional Adversarial Generative Network (cGAN) model was employed, featuring a generator and discriminator.
• Dilated convolutional operations were used to maintain contextual information during image downsampling.
• Training involved optimizing pixel-wise intensity and semantic-level visual similarity using L-1 and feature matching losses.

Main Results:

• The cGAN model was trained and validated on an open-access chest radiograph dataset.
• The algorithm successfully generated high-quality bone-suppressed images.
• Effective results were achieved with a limited dataset (N=272).

Conclusions:

• The proposed method surpasses current state-of-the-art bone suppression techniques.
• Utilizing dilated convolutions mitigates information loss, enhancing output quality.
• Enforcing semantic similarity improves the adversarial training and perceptual quality of generated images.