Anatomy-Decomposed Chest CT Projections Enable Bone Suppression in Chest X-rays Without Paired Data

Forum topic · 小凯 · 2026-09-23

Summary

Researchers propose a digitally reconstructed radiograph (DRR) framework that converts chest CT volumes into pixel-registered paired supervision for anatomy suppression in chest X-rays. A novel bone segmentation algorithm decomposes each CT into bone, non-lung soft-tissue, and lung components, which are projected separately and combined into a synthetic radiograph whose component images sum exactly to the full DRR. Models trained on this data suppress bone or lung components by predicting the target component and recovering the remainder via subtraction, transferring to real radiographs without real paired training data. As an extension, model outputs on real X-rays serve as a target domain for unpaired component-level DRR translation, narrowing the appearance gap while preserving anatomical detail. Experiments on multiple public datasets show that downstream detection gains concentrate on abnormalities with significant bone overlap. Compared with open-source DRR engines on the same CTs, the unmodified DRR achieves comparable realism and label-relevant anatomy preservation, while the translated DRRs achieve the best FID, lung field clarity, and consistency with source CT anatomy. Model and inference code are available at https://huggingface.co/qureaiorg/bone-suppression.

Anatomy-Decomposed Chest CT Projections as Supervision for Bone Suppression in Chest Radiographs

arXiv: 2609.24937 Authors: Mrunmay Angaitkar, Piyush Kumar, Aarjav Satia, Pranav Rao, Ashish Mittal, Manoj Tadepalli, Preetham Putha Posted: 2026-09-21

Overview

Bone overlap can obscure abnormalities in chest radiographs, while scarce paired training data limit supervised bone-suppression approaches. This paper addresses the challenge with a digitally reconstructed radiograph (DRR) framework that converts chest computed tomography (CT) into paired supervision for component suppression.

Method

A novel bone segmentation algorithm decomposes a chest CT into three components: bone, non-lung soft tissue, and lungs, each projected separately.
A weighted combination of the projections yields synthetic radiographs accompanied by pixel-registered component images that sum exactly to the full DRR.
Models trained on this data suppress bone or lung components by predicting the target component and recovering the remainder by subtraction, transferring to real radiographs without real paired training data.
As an extension, model outputs on real radiographs provide a target domain for unpaired component-level DRR translation, narrowing the appearance gap while preserving anatomical detail.

Results

Downstream detection experiments across multiple public datasets demonstrate the utility of bone suppression, with gains concentrated on abnormalities that significantly overlap with bone.
Compared with open-source DRR engines applied to the same CTs, the unmodified DRR achieves comparable realism and preservation of label-relevant anatomy.
The translated DRRs achieve the best FID, lung field clarity, and consistency with source CT anatomy among evaluated methods.

Code

Model and inference code: https://huggingface.co/qureaiorg/bone-suppression