DeBoneDiT: Depth-Driven Conditional Bridge Diffusion Transformers for Bone Suppression in Chest X-Ray Images

Yifei Sun; Fenglei Fan; Junhao Jia; Wenming Deng; Hongxia Xu; Changmiao Wang; Ruiquan Ge

cs.CVOct 14, 2026

DeBoneDiT: Depth-Driven Conditional Bridge Diffusion Transformers for Bone Suppression in Chest X-Ray Images

Authors: Yifei Sun, Fenglei Fan, Junhao Jia, Wenming Deng, Hongxia Xu, Changmiao Wang, Ruiquan Ge

Organizations: Zhejiang University · Hangzhou Dianzi University · City University of Hong Kong · National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital & Shenzhen Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College ·

Abstract

Chest X-Ray (CXR) is a primary modality for diagnosing pulmonary diseases, yet the overlap of bony structures often obscures critical radiographic findings, increasing the risk of diagnostic uncertainty. While Dual-Energy Subtraction (DES) is the clinical benchmark for bone suppression, its widespread adoption is primarily hindered by specialized hardware requirements and increased radiation exposure. Recently, diffusion-based generative models have shown promise for this task, yet they are still constrained by reconstruction-biased latent compression, prior-induced domain gap, and a scarcity of high-quality paired datasets. To address these challenges, we propose DeBoneDiT, a depth-driven conditional bridge diffusion Transformer architecture for efficient high-resolution bone suppression. We first introduce Depth Auto-Encoder (DAE), a structure-aware vision tokenizer for latent compression. By incorporating a multi-level depth feature loss derived from the pretrained Depth Anything V2 encoder, DAE preserves both visual fidelity and structural consistency. Second, we formulate bone suppression as a source-conditioned Brownian bridge diffusion process, leveraging CXRs as both the diffusion prior and structural guidance to narrow the source-to-target domain gap and maintain anatomical plausibility. Furthermore, we have constructed and released SZCH-X-Rays, the largest publicly available high-quality paired dataset for bone suppression to date, comprising 741 pairs of CXR and DES soft tissue images. Extensive experiments and downstream evaluation demonstrate that DeBoneDiT achieves superior bone suppression performance with reduced computational overhead.

Figures & tables

Figure 1: Comparison of directed graphical models of Conditional Diffusion Model (CDM) [ 12 ] , Brownian Bridge Diffusion Model (BBDM) [ 18 ] and DeBoneDiT (Ours). ( $z_{0}$ : latent tissue; $y$ : latent CXR)

Figure 2: Violin plots of data quality using Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) [ 19 ] (

\downarrow

) across two datasets ( S : SZCH-X-Rays, J : JSRT [ 27 ] ). Our SZCH-X-Rays demonstrates higher quality with lower BRISQUE scores for both CXR and soft tissue data.

Figure 3: Overview of the DeBoneDiT architecture. Stage 1 (top): A Depth Auto-Encoder (DAE) is first pretrained to reconstruct images while preserving structural consistency, where a multi-level depth feature loss $\mathcal{L}_{\text{dep}}$ is computed between the input and reconstruction using a pretrained Depth Anything V2 (DAV2) encoder. Stage 2 (bottom): The bone suppression task is formulated as a source-conditioned Brownian bridge diffusion process in the latent space, where a DiT-based network iteratively transforms the latent representation of the source CXR $z_{T}$ into the target soft tissue $z_{0}$ .

Figure 4: Visualization of bone residuals and depth feature discrepancies between a CXR and its corresponding soft tissue image. The depth feature difference map exhibits prominent responses around bony structures, indicating the sensitivity of depth features to skeletal anatomy.

Dataset	Method	Source	LPIPS ( $\downarrow$ )	MSE ( $10^{-3}$ , $\downarrow$ )	PSNR (dB, $\uparrow$ )	SSIM (%, $\uparrow$ )	BSR (%, $\uparrow$ )
SZCH-X-Rays	xU-NetFullSharp	BSPC 25	0.141 $\pm$ 0.032	2.154 $\pm$ 1.544	27.885 $\pm$ 3.418	87.269 $\pm$ 3.685	91.934 $\pm$ 5.482
	Pix2Pix	ISWA 22	0.196 $\pm$ 0.039	1.503 $\pm$ 0.812	28.886 $\pm$ 2.460	83.884 $\pm$ 3.022	92.842 $\pm$ 3.816
	MDA-GAN	TAI 24	0.141 $\pm$ 0.033	1.866 $\pm$ 1.389	28.549 $\pm$ 3.464	87.854 $\pm$ 4.080	92.634 $\pm$ 4.908
	Yang et al.	TMI 25	0.203 $\pm$ 0.039	1.392 $\pm$ 0.592	28.953 $\pm$ 1.855	78.105 $\pm$ 2.797	92.349 $\pm$ 3.153
	BS-Diff	ISBI 24	0.189 $\pm$ 0.038	1.975 $\pm$ 1.137	27.897 $\pm$ 2.921	87.673 $\pm$ 4.005	92.821 $\pm$ 4.980
	BS-LDM	JBHI 25	0.103 $\pm$ 0.016	1.615 $\pm$ 0.933	28.587 $\pm$ 2.435	78.412 $\pm$ 2.861	92.173 $\pm$ 4.314

Table 1: Quantitative comparison of different methods on the SZCH-X-Rays and JSRT datasets. Statistical significance compared to the second-best counterpart is evaluated using the Wilcoxon signed-rank test with Holm-Bonferroni correction applied independently for each dataset (

{}^{*}:p<0.05

{}^{**}:p<0.01

{}^{***}:p<0.001

Figure 5: Visual comparison of different methods on the SZCH-X-Rays and JSRT datasets. The even rows display pseudo-color magnified views of the pixel intensities within the regions indicated by red boxes to highlight subtle textural details. (GT: Ground Truth)

Figure 6: Qualitative comparison of the outputs of DeBoneDiT (top) against the ground truth (bottom) from JSRT. As indicated by the red arrows, the ground truth contains visible acquisition artifacts. In contrast, DeBoneDiT produces anatomically coherent soft tissue structures without reproducing these artifacts. This observation indicates a degree of robustness to imperfections in the reference images and suggests that the model does not simply memorize artifact patterns from the training data.

Method	Scheduler	Timesteps	Params (M)	Time (s)
BS-Diff	DDPM	1000	254.7	110.5
BS-LDM	DDPM	1000	421.3	85.9
GL-LCM	LCM	50	436.9	8.7
DeBoneDiT	BBDM	50	72.8	3.4

Table 2: Quantitative comparison of model complexity and inference latency among different diffusion-based methods.

Dataset	$\mathcal{L}_{\text{dep}}$	LPIPS ( $\downarrow$ )	BRISQUE ( $\downarrow$ )
SZCH-X-Rays	$\times$	0.169 $\pm$ 0.067	26.644 $\pm$ 2.591
	$\checkmark$	0.078 $\pm$ 0.023	22.043 $\pm$ 5.001
JSRT	$\times$	0.061 $\pm$ 0.023	47.003 $\pm$ 0.873
	$\checkmark$	0.032 $\pm$ 0.013	37.144 $\pm$ 3.787

Table 3: Quantitative ablation study on the impact of the depth feature loss

\mathcal{L}_{\text{dep}}

Figure 7: Sensitivity analysis of the depth feature loss weight

\lambda_{\text{dep}}

on JSRT. The results indicate that setting

\lambda_{\text{dep}}=1

attains the minimal perceptual error across (a) CXR reconstruction, (b) soft tissue reconstruction, and (c) bone suppression, demonstrating its optimal balance for the task.

Figure 8: Latent channel correlation analysis of DAE on SZCH-X-Rays. A lower Average Absolute Off-Diagonal Correlation (AAODC) [ 39 ] indicates a higher degree of feature disentanglement.

Dataset	Paradigm	Prior	Condition	LPIPS ( $\downarrow$ )	MSE ( $10^{-3}$ , $\downarrow$ )	PSNR (dB, $\uparrow$ )	BSR (%, $\uparrow$ )
SZCH-X-Rays	CDM	$\mathcal{N}(\mathbf{0},\mathbf{I})$	$y$	0.093 $\pm$ 0.018	2.004 $\pm$ 1.207	27.835 $\pm$ 2.861	92.141 $\pm$ 4.031
	BBDM	$y$	$\times$	0.137 $\pm$ 0.040	1.558 $\pm$ 0.829	28.602 $\pm$ 2.103	86.973 $\pm$ 10.154
	DeBoneDiT	$y$	$y$	0.078 $\pm$ 0.023	1.233 $\pm$ 0.626	29.620 $\pm$ 2.168	93.143 $\pm$ 4.160
JSRT	CDM	$\mathcal{N}(\mathbf{0},\mathbf{I})$	$y$	0.058 $\pm$ 0.024	1.212 $\pm$ 0.881	30.257 $\pm$ 3.138	91.792 $\pm$ 9.406
	BBDM	$y$	$\times$	0.046 $\pm$ 0.015	2.021 $\pm$ 0.972	27.311 $\pm$ 1.665	90.409 $\pm$ 9.532
	DeBoneDiT	$y$	$y$	0.032 $\pm$ 0.013	1.017 $\pm$ 0.983	31.339 $\pm$ 3.280	93.530 $\pm$ 9.296

Table 4: Quantitative ablation study on the impact of different diffusion paradigms, including CDM, BBDM and DeBoneDiT (Ours).

Figure 9: Sampling process comparison of three diffusion paradigms on SZCH-X-Rays. CDM is initialized from Gaussian noise, while BBDM and DeBoneDiT leverage the source CXR as the prior. Notably, DeBoneDiT demonstrates superior detail preservation compared to BBDM.

Figure 10: Loss curves over 2000 epochs. The left and right plots illustrate the training and validation losses, respectively. Incorporating SCI yields superior optimization, converging to a lower loss than the BBDM baseline (w/o SCI).

Figure 11: Quantitative ablation study on the impact of the denoising backbone. Compared to U-Net [ 6 ] and U-ViT [ 3 ] , the DiT-based variants [ 21 ] exhibit superior efficiency and scalability while achieving state-of-the-art bone suppression performance. (

N

: number of DiT blocks)

Figure 12: Quantitative efficiency comparison of various denoising backbones across different resolutions. DiT consistently outperforms U-Net and U-ViT, demonstrating robust scalability and efficiency essential for processing high-resolution medical images.

Figure 13: Visual analysis of the scaling law in bone suppression performance. The red-outlined error maps illustrate the pixel-wise residuals relative to the ground truth, which progressively diminish as the number of DiT blocks $N$ increases, indicating enhanced denoising capability.

Model	Accuracy (%, $\uparrow$ )	Precision ( $\%,\uparrow$ )			Recall (%, $\uparrow$ )			F1-Score (%, $\uparrow$ )
		Pneumonia	COVID-19	Normal	Pneumonia	COVID-19	Normal	Pneumonia	COVID-19	Normal
Original CXR Images
DenseNet-121	90.8 $\pm$ 1.3	97.3 $\pm$ 1.2	96.8 $\pm$ 1.8	81.4 $\pm$ 2.7	86.5 $\pm$ 1.3	89.0 $\pm$ 2.5	96.9 $\pm$ 1.3	91.6 $\pm$ 0.7	92.7 $\pm$ 1.7	88.5 $\pm$ 1.8
GoogLeNet	92.3 $\pm$ 0.8	97.0 $\pm$ 0.8	96.3 $\pm$ 2.0	85.1 $\pm$ 1.0	88.7 $\pm$ 1.1	92.4 $\pm$ 1.9	95.9 $\pm$ 1.1	92.7 $\pm$ 0.3	94.3 $\pm$ 1.5	90.2 $\pm$ 1.0
RegNetY-400MF	88.5 $\pm$ 1.6	90.6 $\pm$ 4.2	97.7 $\pm$ 1.8	80.4 $\pm$ 4.3	90.3 $\pm$ 1.4	83.1 $\pm$ 5.3	92.0 $\pm$ 4.4	90.4 $\pm$ 2.3	89.6 $\pm$ 2.8	85.7 $\pm$ 2.0
ResNet-50	91.2 $\pm$ 0.6	96.8 $\pm$ 3.0	96.2 $\pm$ 1.8	83.1 $\pm$ 1.6	86.7 $\pm$ 2.7	91.0 $\pm$ 2.6	96.0 $\pm$ 2.7	91.4 $\pm$ 1.1	93.5 $\pm$ 1.6	89.1 $\pm$ 0.9

Table 5: Quantitative downstream classification performance on the Asraf dataset using CXRs versus bone-suppressed images, with statistical significance between paired metric values of each backbone under identical 5-fold splits evaluated using the Wilcoxon signed-rank test (

{}^{*}:p<0.05

Medical Imaging

Medical imaging uses computational methods to analyze, reconstruct, and interpret images for diagnosis, treatment planning, and monitoring.

See all 252 papers

Anatomy-Decomposed Chest Computed Tomography (CT) Projections as Scalable Supervision for Bone Suppression in Chest Radiographs
Sep 21, 2026Mrunmay Angaitkar, Piyush Kumar, Aarjav Satia +4Medical Imaging Medical Image Synthesis
Qure.ai, Mumbai, India
DiffusionXRay: A Diffusion and GAN-Based Approach for Enhancing Digitally Reconstructed Chest Radiographs
Mar 2, 2026Aryan Goyal, Ashish Mittal, Pranav Rao +2Image Enhancement Medical Image Synthesis
Indian Institute of Technology Bombay, India · Qure.ai, India
Anatomy-Grounded Weakly Supervised Prompt Tuning for Chest X-ray Latent Diffusion Models
Jun 12, 2025Konstantinos Vilouras, Ilias Stogiannidis, Junyu Yan +2Diffusion Model Alignment Prompt Tuning
School of Engineering, University of Edinburgh · Canon Medical Research Europe Ltd.

Sep 21, 2026J/K move · Enter open · S save

Anatomy-Decomposed Chest Computed Tomography (CT) Projections as Scalable Supervision for Bone Suppression in Chest Radiographs

Mrunmay Angaitkar, Piyush Kumar, Aarjav Satia, Pranav Rao+3

Qure.ai, Mumbai, India

Medical Imaging Medical Image Synthesis

Open paper AskPDF arXiv

Bone overlap can obscure abnormalities in chest radiographs, while scarce paired training data limit supervised bone suppression. We address this challenge with a digitally reconstructed radiograph (DRR) framework that converts chest computed tomography (CT) into paired supervision for component suppression. A novel bone segmentation algorithm enables CT decomposition into bone, non-lung soft-tissue, and lung components, which are projected separately. Their weighted combination yields synthetic radiographs with pixel-registered component images that sum exactly to the full DRR. Models trained on these data suppress bone or lung components by predicting the target component and recovering the remainder by subtraction, transferring to real radiographs without real paired training data. As an extension, their outputs on real radiographs provide target domains for unpaired, component-wise DRR translation, reducing the appearance gap while retaining anatomical details. Across multiple public datasets, downstream detection experiments demonstrate the utility of bone suppression, with gains concentrated on abnormalities with substantial bone overlap. Compared with open-source DRR engines applied to the same CTs, our unmodified DRRs achieve comparable realism and preservation of label-relevant anatomy, while translated DRRs achieve the best Fréchet inception distance (FID), lung-field sharpness, and agreement with source-CT anatomy among the evaluated methods. Models and inference code: https://huggingface.co/qureaiorg/bone-suppression ; Translated projections: https://huggingface.co/datasets/qureaiorg/ct2xr-projections.

Abstract

Medical Imaging

Explore similar work

Anatomy-Decomposed Chest Computed Tomography (CT) Projections as Scalable Supervision for Bone Suppression in Chest Radiographs

DiffusionXRay: A Diffusion and GAN-Based Approach for Enhancing Digitally Reconstructed Chest Radiographs

Anatomy-Grounded Weakly Supervised Prompt Tuning for Chest X-ray Latent Diffusion Models