With the increasing demand for computed tomography (CT) scans, radiology departments are under growing pressure to manage large volumes of imaging data efficiently. Conventional chest X-rays (CXRs) remain a crucial tool in diagnosing various thoracic conditions, but they require separate imaging procedures that add to the clinical workload. Digitally reconstructed radiographs (DRRs) offer a potential alternative, as they are generated directly from CT scans, providing two-dimensional (2D) projections that mimic CXRs. This could allow radiologists to interpret chest images more efficiently, reducing reliance on additional X-ray examinations.
However, for DRRs to be integrated into clinical workflows, their image quality and diagnostic accuracy must be rigorously assessed. This study compared four different DRR generation techniques against CXRs using both quantitative and qualitative evaluations. A recent study published in the Journal of Imaging Informatics in Medicine aimed to determine whether DRRs can provide a viable alternative to conventional CXRs by assessing their diagnostic performance in disease detection and their perceived quality by radiologists.
Comparing DRRs to Conventional CXRs
DRRs are generated using computational techniques that simulate X-ray paths through CT volumes. Two primary projection methods exist: point-source-based and parallel-based projections. Point-source projections mimic the clinical acquisition of CXRs, introducing a slight divergence in the image. Parallel-based projections, by contrast, eliminate divergence, resulting in a different image structure.
The study compared four DRR techniques using a dataset of ultra-low-dose (ULD) chest CTs from 217 patients. For quantitative evaluation, the artificial intelligence model CheXNet, trained to detect 14 thoracic diseases, was applied to both CXRs and DRRs. The model’s performance was measured using area under the curve (AUC) scores. The AUC for CXRs was 0.80, while DRRs produced scores between 0.75 and 0.82, with no statistically significant differences. The SoftMip technique achieved the highest AUC score among DRR methods. These findings suggest that DRRs may be non-inferior to CXRs in automated disease detection.
Radiologists’ Perception of DRR Image Quality
In addition to AI-based evaluation, six radiologists assessed DRRs using a qualitative approach. They rated four key anatomical regions—soft tissue, ossal structures, the mediastinum and the lungs—using a six-point Likert scale, with scores ranging from ‘not diagnostic quality’ to ‘diagnostic quality’. The overall DRR ratings ranged from 3.0 to 3.5, indicating that radiologists found their diagnostic quality to be neutral rather than clearly acceptable or unacceptable.
Several key concerns emerged from this evaluation. One was resolution, as DRRs inherently have lower resolution than CXRs. Standard CXRs are acquired at resolutions exceeding 3000 × 3000 pixels, while DRRs are limited by the CT scan’s resolution, typically around 512 × 512 pixels in the axial plane. Radiologists noted that the reduced resolution made it more challenging to identify fine anatomical details.
Noise was another issue, particularly in soft tissue regions. DRRs derived from ULDCT data exhibited higher noise levels than CXRs, which could impact radiologists’ ability to confidently interpret images. Some radiologists found the increased noise beneficial in certain cases, as it enhanced contrast, but overall, it was viewed as a limiting factor.
Another concern was related to the geometric properties of DRRs. Radiologists noted that point-source projections introduced distortions, particularly at the cranial and caudal edges, where structures overlapped and obscured the view. This over-projection was not present in parallel-based projections, leading to differences in how anatomical features were perceived across different DRR techniques.
Challenges and Potential for Clinical Adoption
Despite these limitations, DRRs offer several advantages that could facilitate their clinical adoption. One of the most significant benefits is efficiency—DRRs can be generated within seconds from existing CT scans without requiring additional X-ray acquisitions. This could be particularly useful in scenarios where CT scans have already been performed, eliminating the need for separate CXR procedures. Additionally, DRRs can be reconstructed from multiple angles, providing flexible imaging options that are not available with standard CXRs.
However, improvements in DRR image quality are necessary before widespread clinical use can be considered. Enhancing resolution, reducing noise and refining the look-and-feel of DRRs to more closely resemble CXRs could increase radiologists’ confidence in their diagnostic reliability. One potential approach is applying advanced denoising techniques or super-resolution algorithms to DRRs to improve clarity without increasing computational complexity.
It is also essential to determine the most suitable use cases for DRRs in clinical practice. While the study focused on DRRs derived from ULDCT scans, further research is needed to explore their applicability across different imaging protocols and patient populations. Understanding how DRRs can best complement existing imaging workflows will be key to their successful integration into routine practice.
The study demonstrated that DRRs can achieve comparable disease detection performance to CXRs in AI-based assessments, suggesting their potential as a diagnostic tool. However, radiologists expressed concerns regarding image quality, particularly in terms of resolution, noise levels and geometric distortions. While DRRs offer advantages in terms of efficiency and flexibility, improvements in image reconstruction techniques are needed to align them more closely with the diagnostic standards of CXRs. Addressing these challenges could enable DRRs to play a meaningful role in optimising radiological workflows, reducing reliance on traditional X-ray imaging and enhancing the efficiency of chest image interpretation.
Source: Journal of Imaging Informatics in Medicine
Image Credit: iStock
