PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.19466v1cs.CV

TL;DR

Bone overlap can hide pulmonary findings in CXRs, and existing diffusion approaches face gaps in structural preservation and paired-data availability. DeBoneDiT combines depth-aware latent compression with a CXR-conditioned Brownian bridge and introduces the 741-pair SZCH-X-Rays dataset. The authors report superior bone suppression with reduced computational overhead, while noting retrospective, single-center evaluation as a limitation.

  • Problem

    CXR bone overlap can obscure pulmonary findings, while diffusion methods are constrained by reconstruction-focused latent compression and scarce high-quality paired datasets.

    Source: §1 · pp. 3–4

  • Method

    DeBoneDiT combines a depth-loss-based autoencoder with a source-conditioned Brownian bridge diffusion process and introduces the SZCH-X-Rays paired dataset.

    Source: §1 · pp. 4–5, §2 · p. 5

  • Results

    On SZCH-X-Rays, DeBoneDiT surpasses the second-best baseline by 0.667 dB in PSNR and achieves an SSIM of 88.204%, with overall superior performance and reduced computational overhead reported.

    Source: §3.3 · p. 10, §4 · p. 16

  • Takeaways & Limitations

    The authors present DeBoneDiT as a potentially practical tool for clinical CXR bone suppression, reporting preservation of pulmonary textures and rapid inference.

    Source: §4 · p. 16

    Evaluation is retrospective and limited to single-center data.

    Source: §3.6 · p. 16

Abstract

from arXiv · show

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.

1 Introduction

Bone overlap can obscure pulmonary findings in CXRs, while the clinical benchmark, DES, is limited by specialized instrumentation, radiation exposure, and motion artifacts. DeBoneDiT addresses gaps in structural preservation, diffusion conditioning, and paired data through a depth-aware tokenizer, a CXR-conditioned bridge process, and the 741-pair SZCH-X-Rays dataset.

1 Introduction

  • CXR bone overlap can obscure pathological details, whereas DES separates bone and soft tissue but is constrained by specialized instrumentation, increased radiation exposure, and motion artifacts.Bone-suppressed images are reported to improve radiologists’ sensitivity and provide cleaner inputs for CAD systems [30].Source: §1 · p. 3
  • Earlier bone-suppression approaches include shallow pixel-level or statistical models, AEs, U-Nets, and GANs, but can lack semantic understanding, require labor-intensive annotations, or lose fine pulmonary detail.Reported problems include over-smoothing, texture loss, and incomplete bone removal.Source: §1 · p. 3
  • Existing diffusion methods often start from Gaussian noise, creating a distribution gap that complicates anatomically plausible mappings and can require extensive iterative denoising.The introduction identifies this alongside structure-unaware latent compression and scarce paired data as a core research challenge.Source: §1 · p. 4
  • DAE uses multi-level depth feature loss from a pretrained Depth Anything V2 encoder to preserve visual fidelity and structural consistency during latent compression.The tokenizer is designed to address reconstruction-focused compression that does not explicitly preserve structures needed to distinguish overlapping bone and soft tissue.Source: §1 · p. 4
  • DeBoneDiT formulates bone suppression as a Brownian bridge diffusion process, using CXRs as the diffusion prior and structural guidance to maintain anatomical plausibility.The authors describe this as a first formulation for bone suppression intended to narrow the source-to-target domain gap.Source: §1 · pp. 4–5
  • SZCH-X-Rays adds 741 paired CXR and DES soft tissue images and is presented as the largest publicly available high-quality paired bone-suppression dataset to date.The dataset addresses a stated scarcity of large-scale, high-quality pairs; the existing public JSRT dataset has 241 paired images with blurring and pronounced artifacts.Source: §1 · pp. 4–5
  • Extensive experiments and downstream evaluation report superior bone suppression with reduced computational overhead for DeBoneDiT.The introduction provides no quantitative values for this comparison.Source: §1 · p. 5

2 Methodology

DeBoneDiT compresses images with a depth-aware tokenizer, then uses source-conditioned Brownian-bridge diffusion to translate CXR latents into soft-tissue latents while preserving anatomical structure.

2 Methodology

  • The two-stage framework combines depth-aware latent compression with source-conditioned Brownian-bridge diffusion for high-resolution bone suppression.The source CXR also provides explicit structural guidance during iterative synthesis.Source: §2 · p. 5

2.1 Depth Auto-Encoder

  • DAE combines reconstruction constraints with multi-level Depth Anything V2 features to preserve visual fidelity and structural consistency during latent compression.Its depth loss compares features across layers of a pretrained depth encoder.Source: §2.1 · pp. 5–7

2.2 Brownian Bridge Diffusion Process

  • The Brownian bridge transitions between target soft-tissue and source CXR latents, with intermediate states interpolating between endpoints and stochastic variance vanishing at both ends.A DiT denoiser predicts added noise to reconstruct the target from intermediate states.Source: §2.2 · p. 7

2.3 Source Condition Injection

  • Source Condition Injection concatenates the noisy latent and source CXR latent at each denoising step, providing explicit anatomical cues to mitigate drift and error accumulation.The source also serves as a boundary constraint during sampling.Source: §2.3 · p. 8

3.1 Datasets and Evaluation metrics

Evaluation uses SZCH-X-Rays and JSRT for bone-suppression performance and Asraf for downstream classification. The section describes dataset preparation and the metrics used for image quality and classification.

3.1.1 Data Preparation

  • SZCH-X-Rays and JSRT evaluate bone-suppression performance, while Asraf is used to benchmark downstream clinical applicability.All images were resized to 1024 × 1024 pixels for experimental consistency.Source: §3.1.1 · p. 9
  • SZCH-X-Rays provides 741 paired posterior-anterior CXRs and DES soft tissue images, split into 592 training, 74 validation, and 75 test pairs.Pairs affected by operational errors, pronounced motion artifacts, severe pleural effusion, or pneumothorax were excluded; the dataset is publicly available under CC BY-NC 4.0.Source: §3.1.1 · p. 9
  • JSRT provides 241 CXR and synthetic soft tissue pairs, split into 192 training, 24 validation, and 25 test pairs.Asraf supplies 6,939 CXRs across pneumonia, COVID-19, and normal classes for downstream evaluation using five-fold cross-validation.Source: §3.1.1 · p. 9

3.1.2 Evaluation Metrics

  • Bone-suppression evaluation uses LPIPS and MSE (lower is better), PSNR, SSIM, and BSR (higher is better).Downstream classification reports accuracy, precision, recall, and F1-score; BRISQUE evaluates generated-image and ground-truth quality.Source: §3.1.2 · p. 9

3.2 Implementation Details

The implementation trains the DAE and DiT denoiser from scratch, then uses 50 diffusion sampling steps for inference. The specified settings include latent dimensions, optimization schedules, and downstream-classifier training.

3.2 Implementation Details

  • The DAE was trained from scratch for 200 epochs with batch size 4, using learning rates of 10−4 for the encoder and decoder and 5 × 10−4 for the discriminator.The Adam optimizer was used.Source: §3.2 · p. 10
  • The DiT denoiser used N = 12 blocks and was trained from scratch for 2000 epochs with batch size 8 using AdamW, with multi-step learning-rate decay initialized at 10−4.EMA and CFG were adopted to ensure model robustness during training.Source: §3.2 · p. 10
  • Diffusion training used 1000 timesteps, while inference used 50 sampling steps to facilitate rapid generation.Source: §3.2 · p. 10
  • Downstream classifiers were fine-tuned for 20 epochs with batch size 16 using ImageNet-1K pretrained weights.Training used Adam, weight decay 10−4, an initial learning rate of 10−3, plateau-based scheduling, and cross-entropy with label smoothing 0.15.Source: §3.2 · p. 10

3.3 Main Results

DeBoneDiT outperforms the evaluated methods on both bone-suppression datasets and preserves fine pulmonary details across them. It also combines fast inference with a compact parameter count.

3.3 Main Results

  • DeBoneDiT outperforms all seven competing methods across five metrics on both SZCH-X-Rays and JSRT.The comparison is reported in Table 1.Source: §3.3 · p. 10
  • 0.667 dB: DeBoneDiT surpasses the second-best baseline in PSNR on SZCH-X-Rays and achieves SSIM of 88.204%.Its LPIPS on SZCH-X-Rays is 0.078; on JSRT, LPIPS is 0.032 versus 0.058 for second-best GL-LCM, alongside BSR of 93.530%.Source: §3.3 · pp. 10–11
  • DeBoneDiT suppresses overlapping bone while preserving fine pulmonary details across SZCH-X-Rays and JSRT.The visual comparisons report over-smoothing by Pix2Pix and BS-Diff, and residual bone shadows from BS-LDM on JSRT.Source: §3.3 · p. 11, Figure 5 · p. 11
  • 3.4 seconds per image: DeBoneDiT uses 72.8 M parameters and is more lightweight than traditional U-Net-based counterparts.Table 2 compares model complexity and inference latency across diffusion-based frameworks.Source: §3.3 · p. 11, Table 2 · p. 11

3.4 Ablation Study

The ablations examine depth-feature supervision, the diffusion paradigm, and denoising backbones. Results favor depth loss and source-conditioned bridge diffusion, while DiT offers a favorable quality-efficiency balance that scales with block count.

3.4 Ablation Study

  • On JSRT, DeBoneDiT produces anatomically coherent soft tissue structures without reproducing visible acquisition artifacts in the ground truth.The comparison suggests robustness to imperfections in reference images and that the model does not simply memorize artifact patterns.Source: Figure 6 · p. 12

3.4.1 Impact of the Depth Feature Loss

  • 0.169 / 0.061 to 0.078 / 0.032: incorporating depth feature loss reduces LPIPS across the reported results.The analysis reports λdep = 1 as the optimal regularization strength for minimizing perceptual distortion and associates lower latent-channel correlation with improved feature disentanglement.Source: Table 3 · p. 12, §3.4.1 · pp. 12–13, Figure 7 · p. 12

3.4.2 Impact of the Diffusion Paradigm

  • 29.620 / 31.339 dB PSNR and 0.078 / 0.032 LPIPS: DeBoneDiT with SCI achieves peak performance against CDM and vanilla BBDM.The comparison reports BBDM LPIPS of 0.137 / 0.046; SCI improves detail retention over BBDM, and Figure 10 reports lower loss with SCI than BBDM without SCI.Source: Table 4 · p. 13, §3.4.2 · p. 13, Figure 9 · p. 14

3.4.3 Impact of the Denoising Backbone

  • DiT achieves a favorable generation-quality and computational-efficiency balance, and increasing its block count consistently improves bone-suppression performance.The backbone comparisons report DiT's efficiency and scalability against U-Net and U-ViT; Figure 13 depicts diminishing residuals as N increases.Source: §3.4.3 · pp. 13–14, Figure 11 · p. 15, Figure 12 · p. 15

3.5 Downstream Evaluation

On the Asraf dataset, DeBoneDiT-generated bone-suppressed images improved downstream classification performance across five representative models and all reported metrics. Evaluation used a frozen model without fine-tuning and five-fold cross-validation.

3.5 Downstream Evaluation

  • The SZCH-X-Rays-pretrained DeBoneDiT model was frozen and applied to Asraf without fine-tuning, followed by standard 5-fold cross-validation.Source: §3.5 · pp. 15–16
  • RegNetY-400MF accuracy rose from 88.5% on original CXRs to 95.5% on DeBoneDiT bone-suppressed images.Source: Table 5 · p. 16, §3.5 · p. 16
  • Across all five classifiers, bone-suppressed images consistently outperformed original CXRs on every evaluation metric, with particularly evident gains in Recall and F1-score.Source: §3.5 · p. 16, Table 5 · p. 16

3.6 Limitations

The authors identify limitations related to occluded lesions, structural supervision, and the scope of clinical evaluation.

3.6 Limitations

  • With 2D projections, pathology completely occluded by hyperdense bone must be reconstructed from learned priors, which may hallucinate or omit extremely subtle lesions.Source: §3.6 · p. 16
  • Depth Anything V2 was trained on natural images, and the authors suggest CT-derived supervision as a future route to anatomy-aware, projection-consistent structural representations.Source: §3.6 · p. 16
  • Evaluation is retrospective and single-center; the authors call for multi-site external validation and prospective clinical studies to strengthen generalizability and clinical-utility assessment.Source: §3.6 · p. 16

4 Conclusion

The paper presents DeBoneDiT, a depth-driven bridge diffusion Transformer for CXR bone suppression, and introduces the paired SZCH-X-Rays dataset. The authors report high-fidelity suppression, rapid inference, and superior performance with reduced computational overhead.

4 Conclusion

  • DeBoneDiT combines structure-aware latent compression, source-conditioned Brownian bridge diffusion, and a DiT-based denoising network for standard CXR bone suppression.Source: §4 · p. 16
  • SZCH-X-Rays is introduced as the largest high-quality paired dataset for bone suppression to date, addressing data scarcity.Source: §4 · p. 16
  • The authors report preserved pulmonary textures, rapid inference, superior performance, and reduced computational overhead, describing the framework as a potential practical tool for clinical CXR bone suppression.Source: §4 · p. 16

Compliance with Ethical Standards

The study used previously obtained, anonymized images retrospectively and received ethics committee approval.

Compliance with Ethical Standards

Loading 2610.19466v1…