Research map / Image generation / Autoregressive and latent generation
BLS-GAN: A Deep Layer Separation Framework for Eliminating Bone Overlap in Conventional Radiographs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BLS-GAN, a GAN framework with a radiography-principle reconstructor and segmentation-based multi-channel supervisor that separates overlapping upper and lower finger bone layers in hand radiographs, with synthetic-image pre-training.
Problem
Bone overlap in conventional radiographs (e.g., MCP joints in rheumatoid arthritis) impedes assessment of individual bone characteristics and automated JSN quantification; standard amodal completion/inpainting does not preserve authentic bone texture.
Approach
Separate bone layers with a GAN whose reconstructor follows X-ray absorption imaging principles (with a learned correction parameter regressed by VGG-18) and whose supervisor is a segmentation network distinguishing overlapped from non-overlapped regions, pre-trained on synthetic overlapped images built from non-overlapping joints.
Key results
- Proposed framework achieves MSE 0.88±0.70 (10^-4), PSNR 41.09±1.72, and FID 1.20±1.33 (10^-2), outperforming MIN-GAN and CMT on MSE and FID.
- Radiological technologists' accuracy on upper and lower bone real-vs-generated sets was about 0.51, near chance, in a visual Turing test.
- Integrating BLS-GAN into deep registration for JSN quantification reduced MSE from 0.0103±0.0133 to 0.0088±0.0118 (P<0.0001).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BLS-GAN instead of MIN-GAN | MCP joint overlap image reconstruction · MCP joint overlap images | MSE (10^-4): 5.63±23.72 -> 0.88±0.70 | Helped |
| Use BLS-GAN instead of MIN-GAN | MCP joint overlap image reconstruction · MCP joint overlap images | SSIM (10^-2): 98.82±3.05 -> 98.37±0.51 | Hurt |
| Use BLS-GAN instead of MIN-GAN | MCP joint overlap image reconstruction · MCP joint overlap images | FID (10^-2): 10.70±50.78 -> 1.20±1.33 | Helped |
| Use BLS-GAN instead of CMT | MCP joint overlap image reconstruction · MCP joint overlap images | MSE (10^-4): 9.26±32.27 -> 0.88±0.70 | Helped |
| Add the correction parameter to the reconstructor (ablation) | BLS-GAN · MCP joint overlap images | MSE (10^-4): 1.72±1.62 -> 0.88±0.70 | Helped |
| Add synthetic-image pre-training (ablation) | BLS-GAN · MCP joint overlap images | MSE (10^-4): 1.05±0.88 -> 0.88±0.70 | Helped |
| Add BLS-GAN bone layer separation before registration | JSN quantification pipeline · JSN progression quantification | MSE: 0.0103 ± 0.0133 -> 0.0088 ± 0.0118 | Helped |
Surprising findings
- The visual Turing test accuracy on joint images was below chance (0.41 overall), indicating synthetic joint images were distinguishable from real ones, though they remained useful for pre-training.
Limitations
- Does not model soft tissue, so soft tissue texture interference remains in overlapped and non-overlapped regions.
- Only evaluated on MCP joints from a single-site RA dataset of 43 patients; generalization to wrist, hip, knee is untested.
- Large overlaps with severe erosion produce unstable generation with sharp edges.
Experimental setup
Datasets: in-house RA hand radiograph MCP joint dataset (1,594 joint images, 672 overlapping) synthetic overlapped images from non-overlapping joints
Benchmarks: in-house RA MCP joint test set
Metrics: MSE SSIM PSNR FID sensitivity specificity accuracy
Baselines: MIN-GAN CMT deep registration without bone layer separation
Builds on: amodal completion VGG Dice loss deep registration for JSN quantification
Closest papers by meaning
- XraySyn: Realistic View Synthesis From a Single Radiograph Through CT Priors (AAAI 2021 · similarity 0.83)
- MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI (NeurIPS 2026 · similarity 0.82)
- RAM-W600: A Multi-Task Wrist Dataset and Benchmark for Rheumatoid Arthritis (NeurIPS 2025 · similarity 0.80)
- Segmenting Maxillofacial Structures in CBCT Volumes (CVPR 2025 · similarity 0.79)
- Learning Low-Rank Feature for Thorax Disease Classification (NeurIPS 2024 · similarity 0.79)
- Large-scale and Fine-grained Vision-language Pre-training for Enhanced CT Image Understanding (ICLR 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.