Accurate characterization of pulmonary nodules and small airways on CT is foundational to early lung cancer detection, risk stratification, and bronchoscopic planning, yet conventional CT spatial resolution has historically limited the assessment of structures smaller than approximately 0.23 to 0.35 mm. Missed or morphologically indeterminate nodules in the submillimeter range may represent early-stage lung cancers with the best potential for cure if detected.
Ultra-high-resolution energy-integrating detector CT (UHR-CT), commercially available since 2017, achieves a spatial resolution of 0.14 mm using smaller detector elements and a smaller x-ray tube focus than conventional CT. Prior studies demonstrated UHR-CT's superior depiction of ground-glass opacities, emphysema, bronchiectasis, and interlobular septal thickening compared to conventional CT, but without histological reference standards to objectively evaluate the accuracy of fine structural representation.
Deep-learning reconstruction (DLR) using convolutional neural networks has emerged as an alternative to iterative reconstruction (IR) for further improving image quality and lesion detectability. However, DLR algorithms are often considered black boxes, raising concerns about whether they accurately represent actual microscopic anatomical structures or introduce artificial alterations to fine morphological features.
Photon-counting detector CT (PCD-CT), a fundamentally different technology available since 2021, achieves 0.11 mm spatial resolution and was previously validated against cadaveric lung histology in the same research group. The current study expands this work to systematically compare UHR-CT across multiple matrix sizes and reconstruction approaches against PCD-CT, with histological images serving as the ground truth reference standard.
Twenty cadaveric human lungs presenting with nodules or ground-glass opacities were scanned under identical conditions on three CT systems: conventional CT (CCT), UHR-CT, and PCD-CT, producing seven distinct image configurations for comparison. The cadaveric specimens were inflated and fixed using the Heitzman method and placed in plastic cases for scanning at radiation doses calibrated to match diagnostic reference levels in thorax phantoms.
UHR-CT used 0.25 mm collimation and super-high-resolution mode. Four UHR-CT reconstructions were generated by varying matrix size and reconstruction method: UHR-512-IR with 0.5 mm slice thickness, UHR-1024-IR with 0.25 mm slice thickness, UHR-2048-IR with 0.25 mm slice thickness (all using Adaptive Iterative Dose Reduction 3D), and UHR-1024-DLR using the Advanced Intelligent Clear-IQ Engine deep learning reconstruction. CCT-512-IR used 0.5 mm slice thickness and served as the baseline reference for energy-integrating detector performance.
PCD-CT used 0.2 mm collimation in ultra-high-resolution mode with 0.2 mm slice thickness, generating PCD-512-IR and PCD-1024-IR configurations. All images were reconstructed with a 350 mm field of view. The primary investigator identified CT slices corresponding to histological sections, and tissue samples were prepared with hematoxylin and eosin staining. Nodule and airway diameters were measured on virtual slide software by the primary investigator and confirmed by a pathologist.
Two experienced chest radiologists independently scored nodules on a five-point scale and airways on a four-point scale by comparing CT images with matched histological sections, blinded to the imaging setting. Discrepancies were adjudicated by a third senior radiologist. Statistical comparisons used the Wilcoxon signed-rank test with Bonferroni correction across nine pre-specified pairs.
Among all energy-integrating detector CT configurations, UHR-2048-IR achieved the highest nodule visualization scores, significantly outperforming CCT-512-IR, UHR-512-IR, and UHR-1024-IR with p less than 0.001 for each comparison. The performance hierarchy for iterative reconstruction images was UHR-2048-IR, then UHR-1024-IR, then UHR-512-IR, then CCT-512-IR, confirming that larger matrix size with thinner slice thickness directly improves nodule delineation.
UHR-1024-DLR showed no statistically significant difference from UHR-2048-IR in overall nodule scores after Bonferroni correction, but for nodules larger than 1000 micrometers, UHR-2048-IR scored significantly better than UHR-1024-DLR (p = 0.003). Deep learning reconstruction showed notable smoothing of irregular margins on some solid nodules, with no UHR-1024-DLR images receiving the maximum score of 5, while both UHR-1024-IR and UHR-2048-IR achieved score-5 results on several solid nodules.
UHR-2048-IR detected nodules with a median barely-detectable diameter of 604 micrometers, compared to 696 micrometers for UHR-1024-IR, demonstrating the gain from the 2048 matrix. PCD-1024-IR outperformed all UHR-CT configurations for nodules larger than 1000 micrometers (p less than or equal to 0.001), reflecting PCD-CT's superior spatial resolution of 0.11 mm versus UHR-CT's 0.14 mm. However, no significant difference was observed between UHR-2048-IR and PCD-1024-IR overall, and UHR-512-IR performed comparably to PCD-512-IR (p = 1.0).
Subgroup analysis showed that for nodules smaller than 500 micrometers and those between 500 and 1000 micrometers, no significant differences were observed across CT series, indicating that the resolution advantages of UHR-CT emerge primarily in the detectability of larger submillimeter and millimeter-scale nodules where morphological detail rather than mere detectability becomes clinically relevant.
For airways, UHR-1024-IR significantly outperformed UHR-512-IR overall (p less than 0.001), confirming that the transition from 512 to 1024 matrix is the critical threshold for bronchiole visualization. Beyond this threshold, no significant differences were observed among UHR-1024-IR, UHR-2048-IR, and UHR-1024-DLR in overall airway scores, indicating diminishing returns from further matrix size increases for airway detection.
UHR-1024-DLR detected the highest number of airways with a score of 4 (lumen and wall both distinctly visualized), and for airways larger than 1000 micrometers, UHR-1024-DLR significantly outperformed PCD-1024-IR (p = 0.005). The noise-reduction capability of deep learning reconstruction appears particularly beneficial for airway lumen-wall boundary delineation, where contrast between the air-filled lumen and soft-tissue airway wall benefits from reduced background noise.
UHR-CT detected barely-discernible airways with median diameters of 665 micrometers (UHR-1024-IR) and 699 micrometers (UHR-2048-IR), demonstrating capability to visualize bronchioles well below the one-millimeter threshold. No significant differences were observed among UHR-CT and PCD-CT configurations for airway visualization at any size category, in contrast to the nodule results where PCD-CT showed clear advantages for larger nodules.
The contrasting behavior of DLR between nodule and airway evaluation provides mechanistic insight: DLR may more accurately reconstruct structures with clear, predictable geometric patterns such as tubular airway lumens, while potentially over-smoothing or altering the irregular, variable margins of solid nodules that deep learning networks may not have been specifically trained to preserve at histological resolution.
The contrasting performance of UHR-1024-DLR across nodule versus airway evaluation reveals an important limitation of deep learning reconstruction that has not previously been established through histological comparison. Irregular margins on solid nodules appeared smoother on DLR images, suggesting that DLR may generate outputs that do not faithfully represent true microscopic lesion morphology, an important consideration for radiological-pathological correlation and for feeding DLR images into radiomics or AI-based analytical pipelines.
This finding is clinically significant because nodule margin irregularity is a critical imaging feature for lung cancer risk stratification. Spiculated or lobulated margins are associated with higher malignancy probability, and if DLR artificially smooths these features, it could lead to underestimation of invasion or nodule aggressiveness in AI-based or radiomics-based classifiers trained on DLR images or applied to DLR output.
The finding that UHR-512-IR performed comparably to PCD-512-IR at the same matrix size has practical clinical implications. PCD-CT systems are substantially more expensive than UHR-CT, and if the performance gap disappears at standard clinical matrix sizes, UHR-CT represents a cost-effective path to resolution improvement over conventional CT without requiring the capital investment of PCD-CT technology.
The study's cadaveric design eliminated motion artifacts and cardiac pulsation that would reduce image quality in vivo, meaning real clinical images will likely show somewhat lower performance than the results reported here. Additional limitations include evaluation of only axial images matching histological slice positions, non-standardized slice thickness between CT systems, and the lack of a DLR setting for the 512 matrix or a DLR option for PCD-CT.
The ability of UHR-CT to detect and morphologically characterize submillimeter nodules and bronchioles smaller than one millimeter in diameter opens new possibilities for earlier lung cancer detection, particularly for adenocarcinoma subtypes that may present as small ground-glass or part-solid nodules with prognostically important morphological features. Accurate margin evaluation of small nodules, rather than mere detection, may be the more clinically impactful application of ultra-high resolution.
Airway visualization improvements have direct implications for bronchoscopy planning and airway disease evaluation. The ability to detect bronchioles approaching 700 micrometers in diameter could improve navigation bronchoscopy route planning for peripheral lesions accessible only through small distal airways, potentially expanding the reach of minimally invasive sampling beyond what is currently guidable with standard CT bronchograms.
A prior radiomics study using 1024-matrix UHR-CT images demonstrated utility in predicting solid and micropapillary components of lung adenocarcinoma, histological features that confer worse prognosis and influence surgical extent decisions. Higher-resolution imaging that better captures these morphological correlates could therefore improve preoperative adenocarcinoma subtype characterization non-invasively.
The comparison between UHR-CT and PCD-CT at matching matrix sizes provides practical guidance for institutions deciding on technology investments: for standard clinical matrix sizes (512), the two technologies perform comparably, while PCD-CT retains clear advantages for high-matrix nodule assessment of larger lesions, and UHR-CT with DLR exceeds PCD-CT for airway visualization in larger bronchioles.
UHR-2048-IR achieved the best overall nodule visualization among all energy-integrating detector CT configurations tested, with a histologically validated ability to detect nodules and airways at median barely-detectable diameters of 604 micrometers and 699 micrometers respectively. These results confirm that incremental matrix size increases on UHR-CT directly translate to improved morphological fidelity verified against histological ground truth.
The finding that UHR-512-IR performs comparably to PCD-512-IR at matched matrix sizes repositions UHR-CT as a clinically accessible high-resolution solution that may provide PCD-CT-level performance for the matrix sizes most commonly used in current clinical practice, without the substantial additional cost of photon-counting detector systems.
Deep learning reconstruction improves airway visualization and reduces image noise, but introduces concerning morphological smoothing of irregular solid nodule margins at the histological level. This DLR-specific limitation was not previously documented with histological validation and represents an important caution for radiologists, AI developers, and radiomics researchers using DLR images to characterize malignant nodule features.
Future studies should evaluate these CT configurations in in vivo lung CT acquisitions that include motion artifacts and cardiorespiratory influences, investigate the impact of high-resolution imaging on AI and radiomics analysis of pulmonary nodules, and assess UHR-CT with DLR at the 512 matrix size to fill the remaining gap in the reconstruction configuration space tested here.