Bone suppression on pediatric chest radiographs via a deep learning-based cascade model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Radiology

Background:

• Bone suppression images (BSIs) enhance pulmonary disease diagnosis on chest X-rays (CXRs).
• Conventional methods like dual-energy subtraction (DES) are unsuitable for pediatric patients due to radiation concerns.
• Existing deep learning models often rely on DES-derived data, limiting their application in pediatrics.

Purpose of the Study:

• To develop a novel, radiation-free bone suppression method specifically for pediatric chest X-rays.
• To create accurate bone-suppressed images for improved visualization of lung abnormalities in children.
• To establish a deep learning framework for pediatric bone suppression without relying on DES.

Main Methods:

• A 2-channel contrastive unpaired image-to-image translation network was trained using adult digitally reconstructed radiographs (DRRs) to generate pseudo-CXRs.
• This model was then applied to 129 pediatric DRRs to create paired training data of pseudo-pediatric CXRs.
• A U-Net model was trained on this pediatric data to perform bone suppression on pediatric CXRs.

Main Results:

• The developed bone suppression model achieved a score of 3.31 ± 0.48 from an expert radiologist.
• Evaluation metrics demonstrated effective bone removal with preserved soft tissue detail.
• Subtle residual bone shadows were observed, but overall bone subtraction was homogeneous.

Conclusions:

• The novel method successfully generates bone-suppressed pediatric chest X-rays.
• Preservation of soft-tissue pixel intensity and effective bone subtraction aid in detecting early pulmonary disease.
• This radiation-free approach offers a safe and effective tool for pediatric lung imaging analysis.