BLS-GAN: A Deep Layer Separation Framework for Eliminating Bone Overlap in Conventional Radiographs

Haolin Wang, Yafei Ou, Prasoon Ambalathankandy, Gen Ota, Pengyu Dai, Masayuki Ikebe, Kenji Suzuki, Tamotsu KamishimaAAAI 2025 · DOI

TL;DR

This paper introduces BLS-GAN, a novel deep learning framework that tackles the pervasive problem of bone overlap in conventional radiographs, a common issue that hinders accurate diagnosis of musculoskeletal diseases like rheumatoid arthritis. By effectively separating overlapping bone layers, BLS-GAN generates high-quality bone images that improve the performance of downstream diagnostic tasks, such as joint space narrowing quantification.

What Problem Does This Paper Address?

Conventional radiography is a cornerstone of musculoskeletal (MSK) imaging due to its accessibility and cost-effectiveness. However, a significant limitation is the phenomenon of bone overlap, where different bones or parts of the same bone obscure each other in the 2D projection. This overlap creates a textural mixture that makes it difficult for both human experts and automated systems to accurately assess bone characteristics, leading to challenges in diagnosing, monitoring, and prognosticating MSK diseases. For instance, in rheumatoid arthritis (RA), bone erosion and joint space narrowing (JSN) are critical indicators, but bone overlap can significantly complicate their assessment. Existing methods like amodal completion or object removal are not well-suited for this specific problem due to the unique imaging principles of radiography and the complex textural nature of bone.

Methods

The core of the proposed solution is the Bone Layer Separation GAN (BLS-GAN) framework. It comprises three main components:

  • Layer Image Generator: This network takes the original radiograph and bone masks as input and generates separate images for each bone layer (e.g., upper and lower bones of a joint). The generator is designed to eliminate overlapping regions.
  • Segmentation-based Multi-channel Supervisor: Unlike traditional GAN discriminators, this supervisor performs pixel-level discrimination. It outputs masks that identify overlapped vs. non-overlapped regions and assess the authenticity of the generated images, providing more granular supervision to the generator.
  • Radiography Imaging Principles-based Reconstructor: This component is crucial for adhering to the physics of X-ray imaging. It reconstructs images based on the generated layer images and a learned correction parameter. This reconstructor aims to mitigate issues arising from soft tissue interference in overlapped regions, which can degrade reconstruction quality. A VGG-18 network is used for the correction parameter regression.

A key innovation is the pre-training strategy using synthetic images. These synthetic images are created by artificially introducing overlaps into non-overlapped bone images, allowing for the generation of ground truth data for training. This pre-training step significantly enhances the stability of the training process and the quality of the generated images.

Datasets

The study utilizes a newly curated dataset specifically for this challenging scenario. This dataset consists of 168 posteroanterior (PA) radiographs of the hand from 43 patients with RA, focusing on metacarpophalangeal (MCP) joints. The dataset includes joint images and corresponding mask annotations for the upper and lower bones. A significant portion (672 images) exhibits bone overlap. The dataset was manually annotated by a radiological laboratory assistant and reviewed by radiologists.

Key Insights and Ideas

  • Bone overlap is a fundamental challenge in radiography: The paper convincingly argues that bone overlap is not just a minor inconvenience but a significant impediment to accurate MSK diagnosis and automated analysis, particularly for diseases like RA.
  • Radiography physics matters: Unlike natural image inpainting, simply filling in missing pixels isn't enough. The BLS-GAN framework's reconstructor explicitly incorporates radiography principles, which is vital for generating realistic bone textures.
  • Supervision is key: The multi-channel supervisor, which distinguishes between overlapped and non-overlapped regions, provides more targeted feedback to the generator than a standard GAN discriminator, leading to better separation.
  • Synthetic data for pre-training is a game-changer: The use of synthetically generated overlapped images for initial training dramatically improves the stability and performance of the model, especially when dealing with limited real-world data for specific scenarios.
  • Improved downstream task performance: The ultimate validation comes from demonstrating that BLS-GAN's output directly benefits clinical applications, such as improving the accuracy and robustness of JSN quantification.

Key Results

  • Quantitative Evaluation: BLS-GAN achieved superior performance across various metrics (MSE, SSIM, PSNR, FID) compared to other amodal completion methods, particularly in reducing MSE and FID. For example, the MSE was significantly lower (0.0088 ± 0.0118) with BLS-GAN compared to a pipeline without it (0.0103 ± 0.0133) in JSN quantification.
  • Qualitative Evaluation: Visual Turing tests conducted with radiological technologists showed that while distinguishing real from generated joint images was challenging (around 0.5 accuracy for upper/lower bone layers), the generated images were highly realistic and preserved bone texture. Figure 4 visually demonstrates the successful separation of bone layers.
  • Robustness: The framework showed improved stability and accuracy in JSN quantification, especially for overlap sizes less than 1000 pixels, and consistently outperformed methods without bone layer separation even at larger overlap sizes.
  • Ablation Studies: Both the correction parameter in the reconstructor and the synthetic image pre-training were shown to significantly improve reconstruction quality, stability, and reduce brightness amplification in overlapped areas.

Limitations and Future Work

  • Soft Tissue Interference: The current framework primarily focuses on bone layer extraction and does not fully address the interference of soft tissue textures, especially in non-overlapped regions. While a correction parameter helps, it doesn't completely remove soft tissue texture interference.
  • Lack of Positive Samples for Soft Tissue: Future work aims to accurately generate and differentiate soft tissue regions from bone layers. This is challenging due to the absence of "positive samples" (images without bones) for supervision in the context of soft tissue generation.
  • Generalization to Other Joints: While the framework is designed to extend to complex joints like the wrist and hip, the current experiments focus on MCP joints. Further validation on a wider range of joints with varying overlap complexities is needed.

Summary updated

BibTeX
@inproceedings{wang2025bls,
  author = {Haolin Wang and Yafei Ou and Prasoon Ambalathankandy and Gen Ota and Pengyu Dai and Masayuki Ikebe and Kenji Suzuki and Tamotsu Kamishima},
  title = {{BLS-GAN: A Deep Layer Separation Framework for Eliminating Bone Overlap in Conventional Radiographs}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i7.32826},
}
Figure