Anatomy-Decomposed Chest Computed Tomography (CT) Projections as Scalable Supervision for Bone Suppression in Chest Radiographs
CT decomposition projections enable scalable bone suppression in chest X-rays, improving abnormality detection accuracy.
Key Findings
Methodology
This study presents a DRR generation framework based on CT anatomical decomposition, converting chest CT into bone, non-lung soft-tissue, and lung components, which are projected separately and combined to create synthetic radiographs. Models trained on these data suppress bone or lung components by predicting the target component and recovering the remainder by subtraction.
Key Results
- On TBX11K and Node21 datasets, bone suppression models significantly improved detection accuracy, achieving an FID of 8.2 and 2.6-3.8 times lung-field sharpness.
- Unmodified DRRs achieved comparable realism and anatomical preservation compared to open-source DRR engines.
- Translated DRRs achieved the best FID, lung-field sharpness, and agreement with source-CT anatomy.
Significance
This study provides scalable supervision for bone suppression using CT decomposition projections, addressing the scarcity of paired training data in traditional methods. Its utility is demonstrated across multiple public datasets, particularly in detecting abnormalities with significant bone overlap.
Technical Contribution
The study introduces a novel bone segmentation algorithm and multi-threshold bone segmentation technique, combined with a polychromatic HU-to-attenuation model, achieving precise structural decomposition and projection. Compared to existing methods, it offers better anatomical preservation and translation performance.
Novelty
This is the first to use CT decomposition projections for chest X-ray bone suppression, providing a solution without paired training data, significantly reducing the appearance gap between synthetic and real radiographs.
Limitations
- The model may encounter errors when handling complex anatomical structures, especially in atypical lesion areas.
- Requires high-quality CT data as input, which may limit its application in resource-constrained environments.
Future Work
Future research could explore more complex anatomical decompositions and more efficient unpaired translation methods to further enhance model adaptability and accuracy.
AI Executive Summary
Chest X-rays are a common diagnostic tool, but bone overlap can obscure potential abnormalities, affecting diagnostic accuracy. Existing bone suppression methods are limited by the scarcity of paired training data, hindering their scalability. This paper proposes a digitally reconstructed radiograph (DRR) generation framework based on CT anatomical decomposition, converting chest CT into bone, non-lung soft-tissue, and lung components, which are projected separately and combined to create synthetic radiographs. Models trained on these data suppress bone or lung components by predicting the target component and recovering the remainder by subtraction. Experimental results demonstrate significant improvements in abnormality detection accuracy across multiple public datasets, particularly in cases with significant bone overlap. Compared to existing open-source DRR engines, this method excels in realism and anatomical preservation. Future research could explore more complex anatomical decompositions and more efficient unpaired translation methods to further enhance model adaptability and accuracy.
Deep Analysis
Background
Chest X-rays are among the most frequently performed diagnostic imaging examinations, but their two-dimensional projection nature leads to bone overlap, affecting lesion detection. Traditional dual-energy subtraction techniques can separate bone from soft tissue but require specialized equipment, making data scarce and susceptible to motion and noise artifacts. Chest CT offers an alternative by resolving superimposed structures in three dimensions, enabling digitally reconstructed radiograph (DRR) generation.
Core Problem
Bone overlap in chest X-rays can obscure potential abnormalities, affecting diagnostic accuracy. Existing bone suppression methods are limited by the scarcity of paired training data, hindering their scalability. How to achieve effective bone suppression without paired data is a pressing issue.
Innovation
This paper proposes a DRR generation framework based on CT anatomical decomposition, converting chest CT into bone, non-lung soft-tissue, and lung components, which are projected separately and combined to create synthetic radiographs. This method is the first to achieve bone suppression without paired training data, significantly reducing the appearance gap between synthetic and real radiographs.
Methodology
- �� Decompose chest CT into bone, non-lung soft-tissue, and lung components. • Use a multi-threshold bone segmentation algorithm for decomposition. • Project each component separately to create synthetic radiographs. • Train models on these data to suppress bone or lung components.
Experiments
Experiments were conducted on TBX11K, Node21, and VinDr-CXR datasets to evaluate the detection performance of bone suppression models. Fréchet inception distance (FID) and lung-field sharpness were used as evaluation metrics. Comparisons with open-source DRR engines validated the realism and anatomical preservation of the method.
Results
Experimental results show that bone suppression models significantly improved abnormality detection accuracy across multiple datasets, particularly in cases with significant bone overlap. Compared to existing open-source DRR engines, this method excels in realism and anatomical preservation.
Applications
This method can improve the accuracy of abnormality detection in chest X-rays, particularly in cases with significant bone overlap. It has significant implications for medical image analysis and computer-aided diagnosis systems.
Limitations & Outlook
The model may encounter errors when handling complex anatomical structures, especially in atypical lesion areas. Requires high-quality CT data as input, which may limit its application in resource-constrained environments.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen, and a CT scan is like separating all the ingredients: bones, meat, and vegetables. Then, you cook them in different pots and finally combine them into one dish. This process is like decomposing a CT into different components, projecting them separately, and then combining them into a complete radiograph. The benefit is that you can see each ingredient (or component) more clearly without being obscured by bones (or other components).
ELI14 Explained like you're 14
Imagine you're playing a game with many levels, each with different monsters. A CT scan is like separating these levels so you can see each monster clearly. Then, you use different weapons to attack them separately and finally combine them for a complete victory. This process is like decomposing a CT into different components, projecting them separately, and then combining them into a complete radiograph. The benefit is that you can see each monster (or component) more clearly without being obscured by other monsters (or components).
Glossary
Digitally Reconstructed Radiograph (DRR)
A synthetic radiograph generated from CT data to simulate real X-rays.
Used in this paper to generate training data for bone suppression.
Bone Suppression
A method to remove bone overlap in X-rays to improve lesion detection accuracy.
Achieved by predicting the target component and recovering the remainder by subtraction.
Fréchet Inception Distance (FID)
A metric for evaluating the quality of generated images, with lower values indicating higher quality.
Used to assess the realism of generated DRRs.
Unpaired Translation
An image translation technique without paired data, preserving content while changing style.
Used to translate generated component projections to the real radiograph domain.
Multi-threshold Bone Segmentation
An algorithm that segments bones using multiple thresholds to ensure precise bone segmentation.
Used in the CT decomposition step.
Open Questions Unanswered questions from this research
- 1 How to apply this method in resource-constrained environments, especially with limited high-quality CT data.
- 2 How to further enhance model adaptability and accuracy in complex anatomical structures.
Applications
Immediate Applications
Medical Image Analysis
Improve the accuracy of abnormality detection in chest X-rays, particularly in cases with significant bone overlap.
Long-term Vision
Computer-aided Diagnosis
Enhance the performance of computer-aided diagnosis systems through improved bone suppression techniques.
Abstract
Bone overlap can obscure abnormalities in chest radiographs, while scarce paired training data limit supervised bone suppression. We address this challenge with a digitally reconstructed radiograph (DRR) framework that converts chest computed tomography (CT) into paired supervision for component suppression. A novel bone segmentation algorithm enables CT decomposition into bone, non-lung soft-tissue, and lung components, which are projected separately. Their weighted combination yields synthetic radiographs with pixel-registered component images that sum exactly to the full DRR. Models trained on these 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, their outputs on real radiographs provide target domains for unpaired, component-wise DRR translation, reducing the appearance gap while retaining anatomical details. Across multiple public datasets, downstream detection experiments demonstrate the utility of bone suppression, with gains concentrated on abnormalities with substantial bone overlap. Compared with open-source DRR engines applied to the same CTs, our unmodified DRRs achieve comparable realism and preservation of label-relevant anatomy, while translated DRRs achieve the best Fréchet inception distance (FID), lung-field sharpness, and agreement with source-CT anatomy among the evaluated methods. Models and inference code: https://huggingface.co/qureaiorg/bone-suppression; Translated projections: https://huggingface.co/datasets/qureaiorg/ct2xr-projections.