[Paper Review] Learning Bone Suppression from Dual Energy Chest X-rays using Adversarial Networks

Dong Yul Oh, Il Dong Yun|arXiv (Cornell University)|Nov 5, 2018

Medical Imaging Techniques and Applications27 references14 citations

TL;DR

This paper proposes a deep generative model using conditional generative adversarial networks (cGANs) with Haar 2D wavelet decomposition to predict bone-suppressed chest X-rays from single-energy images, trained on a limited dataset of dual-energy X-rays. The method achieves state-of-the-art performance by combining adversarial training for image sharpness and wavelet-based frequency guidance for faster, more accurate convergence, outperforming existing methods in bone suppression quality.

ABSTRACT

Suppressing bones on chest X-rays such as ribs and clavicle is often expected to improve pathologies classification. These bones can interfere with a broad range of diagnostic tasks on pulmonary disease except for musculoskeletal system. Current conventional method for acquisition of bone suppressed X-rays is dual energy imaging, which captures two radiographs at a very short interval with different energy levels; however, the patient is exposed to radiation twice and the artifacts arise due to heartbeats between two shots. In this paper, we introduce a deep generative model trained to predict bone suppressed images on single energy chest X-rays, analyzing a finite set of previously acquired dual energy chest X-rays. Since the relatively small amount of data is available, such approach relies on the methodology maximizing the data utilization. Here we integrate the following two approaches. First, we use a conditional generative adversarial network that complements the traditional regression method minimizing the pairwise image difference. Second, we use Haar 2D wavelet decomposition to offer a perceptual guideline in frequency details to allow the model to converge quickly and efficiently. As a result, we achieve state-of-the-art performance on bone suppression as compared to the existing approaches with dual energy chest X-rays.

Motivation & Objective

• To address the clinical limitations of dual-energy subtraction (DES), including doubled radiation exposure and motion artifacts from heartbeats.
• To develop a deep learning model that predicts bone-suppressed images from single-energy chest X-rays using only a finite set of previously acquired dual-energy scans.
• To improve image quality in bone suppression by replacing pixel-wise loss minimization with adversarial training to preserve lesion details.
• To accelerate model convergence and enhance perceptual quality by incorporating Haar 2D wavelet decomposition as a frequency-domain perceptual guide.
• To establish a novel, data-efficient deep generative framework for bone suppression that outperforms existing methods.

Proposed method

• Uses a conditional generative adversarial network (cGAN) to learn the mapping from single-energy chest X-rays to bone-suppressed images, replacing traditional regression with adversarial loss.
• Incorporates Haar 2D wavelet decomposition of input images to provide frequency-domain guidance, helping the generator learn high-frequency bone structures and improve convergence.
• Applies a perceptual loss based on wavelet coefficients to guide the generator in preserving fine anatomical details while suppressing bones.
• Trains the generator to minimize adversarial loss (fooling the discriminator) while also minimizing pixel-wise reconstruction loss (L1) on the wavelet-transformed features.
• Uses a discriminator that distinguishes between real bone-suppressed images (from DES) and generated ones, encouraging realistic texture and structure.
• Employs histogram matching between generated and original images to maintain consistent intensity distribution for visual consistency.

Experimental results

Research questions

• RQ1 Can adversarial training improve bone suppression quality compared to standard pixel-wise regression in low-data regimes?
• RQ2 Does incorporating Haar 2D wavelet decomposition as a perceptual prior enhance the convergence and fidelity of bone suppression models?
• RQ3 Can a deep generative model trained on limited dual-energy X-ray pairs generalize well to single-energy input and outperform existing methods?
• RQ4 How does the combination of adversarial training and wavelet-based guidance affect the preservation of pathological features and image sharpness?
• RQ5 Is the proposed method clinically viable by reducing artifacts and radiation exposure compared to dual-energy subtraction?

Key findings

• The proposed method achieves state-of-the-art bone suppression performance, outperforming existing approaches in both quantitative metrics and visual quality.
• The ablation study confirms that combining cGAN with Haar wavelet decomposition yields the best results, with the CNN + GAN + Haar configuration (ours) significantly outperforming CNN-only and GAN-only variants.
• Adversarial training alone improves sharpness and reduces blur compared to standard L1-based regression, but performance is further enhanced when combined with wavelet guidance.
• The use of Haar wavelet decomposition reduces training time and improves convergence speed, especially when paired with adversarial loss.
• The model successfully suppresses ribs and clavicles while preserving soft-tissue structures and pathological features, as validated through visual and quantitative comparisons with DES ground truth.
• The method effectively mitigates motion artifacts common in dual-energy imaging by generating bone-suppressed images from a single acquisition.