DeBoneDiT: Depth-Driven Conditional Bridge Diffusion Transformers for Bone Suppression in Chest X-Ray Images
Bone suppression in chest X-rays removes overlapping ribs and clavicles that hide lung pathology, but existing methods suffer from reconstruction-biased latent compression, domain gaps, and very small paired datasets. DeBoneDiT tackles this by combining a depth-supervised auto-encoder with a source-conditioned Brownian bridge diffusion Transformer, and it releases SZCH-X-Rays, a 741-pair high-resolution dataset. The goal is to provide an efficient, software-only alternative to hardware-dependent Dual-Energy Subtraction.
DeBoneDiT is a technically solid two-stage pipeline that sets new pixel-wise and perceptual benchmarks on both SZCH-X-Rays and JSRT while cutting inference latency to 3.4 s and the parameter count to 72.8 M. However, the training data remain tiny by modern generative standards, the depth regularizer is a natural-image pretrained model, and the downstream clinical gains, while encouraging, lack the controls needed to isolate the method from simpler preprocessing baselines.
The depth-driven auto-encoder is a compelling addition: the multi-level depth feature loss yields large LPIPS improvements (0.169→0.078 on SZCH and 0.061→0.032 on JSRT), indicating that depth features meaningfully regularize structural fidelity. The use of a Brownian bridge rather than Gaussian noise initialization is also theoretically well-motivated for image-to-image translation, and the explicit source-conditioning mechanism produces sharper anatomical detail than the plain bridge baseline.
The foremost limitation is data scale and generalizability: 741 training pairs, while the largest public set in this niche, is tiny for a latent diffusion model, and the authors concede that evaluation is retrospective and single-center. When pathologies are fully hidden by dense bone, the model must hallucinate soft-tissue content, risking missed subtle lesions. Moreover, the depth supervision comes from Depth Anything V2, a model pretrained on natural images; the paper itself notes that CT-derived structural priors would be preferable. Finally, reference-based metrics are suspect because the JSRT soft-tissue ground truth is algorithmically synthesized and visibly artifact-laden, yet the method is largely evaluated against it.
The reported metrics support improvements over GAN and diffusion baselines, but the comparisons lean heavily on reference-based scores (MSE, PSNR, SSIM, BSR) that are fragile when the reference itself is flawed. On JSRT, DeBoneDiT outperforms the ground truth in perceived artifact removal, which paradoxically undermines those same reference metrics. The downstream classification experiment on Asraf shows consistent gains across five backbones, yet it only contrasts original CXRs against DeBoneDiT outputs; it does not test whether competing suppression methods would confer similar benefits, nor does it control for intensity rescaling or other preprocessing differences.
The authors disclose many training hyperparameters—learning rates, batch sizes, codebook size $K=1024$, diffusion steps $T=1000$ with 50 sampling steps, and the use of EMA and CFG—yet several architectural details required to reproduce the DiT backbone (e.g. hidden dimension, patch size, attention heads) are absent from the main text. Code has not been released, and while the SZCH-X-Rays dataset is publicly available, full reproduction would require re-implementing both DAE and the custom bridge sampler from scratch. The sensitivity analysis for $\lambda_{\text{dep}}$ and the scaling curves for the number of DiT blocks are helpful, but they do not substitute for open-source implementation.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.