Chest X-ray Bone Suppression for Improving Classification of Tuberculosis-Consistent Findings

Area of Science:

• Medical Imaging and Artificial Intelligence
• Radiology and Deep Learning Applications

Background:

• Chest X-rays (CXRs) are crucial for diagnosing cardiopulmonary conditions but can be hindered by bony structures.
• Obscuring bony structures like ribs and clavicles in CXRs can lead to diagnostic errors and misinterpretations.
• Existing deep learning (DL) workflows may be impacted by these limitations, affecting the detection of subtle abnormalities.

Purpose of the Study:

• To develop and evaluate a DL-based bone suppression model for frontal CXRs.
• To improve the accuracy of radiological interpretations, particularly for detecting tuberculosis (TB) manifestations.
• To assess the impact of bone-suppressed images on the performance of DL models for TB classification.

Main Methods:

• Trained and optimized several DL bone suppression models using a combined loss function.
• Evaluated model performance using metrics like MAE, PSNR, SSIM, and MS-SSIM in a cross-institutional setting.
• Fine-tuned a pre-trained VGG-16 model on bone-suppressed and non-bone-suppressed TB CXR datasets (Shenzhen and Montgomery) for TB classification.

Main Results:

• The best bone suppression model (ResNet-BS) achieved high performance (PSNR = 34.0678; MS-SSIM = 0.9828).
• DL models trained on bone-suppressed CXRs demonstrated significantly higher accuracy (AUC: 0.9535-0.9635) compared to those trained on non-bone-suppressed CXRs (AUC: 0.8991-0.8567).
• Bone suppression enhanced model sensitivity towards TB classification and improved the detection of TB-consistent findings.

Conclusions:

• DL-based bone suppression effectively removes obscuring bony structures in CXRs.
• Bone-suppressed CXRs significantly improve the performance of DL models in detecting tuberculosis.
• This approach holds promise for reducing diagnostic errors and enhancing AI-assisted radiological interpretation.