hugging-apps/bone-suppression-chest-xray

Bone & lung-component suppression for chest radiographs

Gradio demo for the suppression models of

Anatomy-Decomposed Chest Computed Tomography (CT) Projections as Scalable Supervision for Bone Suppression in Chest Radiographs — Angaitkar, Kumar, Satia, Rao, Mittal, Tadepalli, Putha (arXiv:2609.24937, 2026; Qure.ai).

Weights: `qureaiorg/bone-suppression` — two TorchScript traces, loaded and run verbatim as in the authors' reference script (suppress.py); the preprocessing contract (1024×1024 area resize, per-image min–max normalisation to [0,1], single channel, bone bright) is reproduced exactly.

What it does

A frontal (PA/AP) chest radiograph is decomposed into four images by applying the two models in sequence, as the paper does:

full radiograph --[bone model]--> bone image        ; soft tissue   = full − bone
soft tissue     --[lung model]--> lung component    ; non-lung soft = soft − lung

Each model predicts one component; the complement is recovered by subtraction. The lung model consumes the floating-point residual soft = full − bone exactly as computed, with no second normalisation — that is the trained inference pipeline.

Notes on the outputs

The predictions live on the normalised input's [0,1] scale, so the raw component images look dark (in the authors' example the bone image peaks near 62/255). By default the panels are contrast-stretched (each clipped to its 0.5–99.5 percentile range) for viewing — the paper's decomposition figure does the same. Switch Output scale to Raw to get the exact files the reference script writes.

The models require a bone-bright radiograph, as displayed clinically; feeding an inverted image gives meaningless output. Auto-invert if bones appear dark detects and corrects this (a scale-invariant comparison of the mediastinum against the lung fields).

Limitations

Trained on synthetic supervision only (per-structure projections rendered from chest CT); no dual-energy pairs and no paired real radiographs. Evaluated on adult frontal radiographs from public datasets (TBX11K, Node21, VinDr-CXR, JSRT) — not evaluated on paediatric, lateral, or portable/supine images. The lung-component output is an experimental model-derived estimate of vessels and other intrapulmonary structure, with no standalone downstream validation in the paper.

This is research software, not a medical device, and is not for diagnostic or clinical use.

Licences

• —Weights (weights/*.ts) — CC BY-NC-SA 4.0, non-commercial research and educational use only (see weights/LICENSE-WEIGHTS.txt in the model repo).
• —Reference code (suppress.py, config.json) — Apache-2.0.
• —Example radiographs in examples/ — from the authors' own `qureaiorg/ct2xr-projections` dataset (CC BY-NC-SA 4.0). They are synthetic CT-derived projections of CT-RATE volumes, redistributed here under the same share-alike terms with attribution to Qure.ai.

Hardware

ZeroGPU (zero-a10g), single @spaces.GPU(duration=...) call covering both models.