Cynthia H. McCollough

Objective: Accurate lung volumetry is essential for donor-recipient size matching in lung transplantation. This study evaluated the impact of radiation dose reduction on the robustness of AI-based whole-lung segmentation, 5-lobe segmentation, and automated tissue classification. Methods: Fifty cases from the Cancer Image Archive (TCIA) Low-dose CT Image and Projection Data set were analyzed. A reduced-dose data set at 10% of the routine dose for each case was generated using validated projection-domain noise insertion. Whole-lung and 5-lobe segmentation were performed using the Lung CT Segmenter module with the lungmask deep learning model in 3D Slicer. Segmentation accuracy was quantified using the Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), and maximum Hausdorff distance (HDmax). Automated tissue classification and volumetry were performed using the Lung CT Analyzer module. Results: Whole-lung segmentation demonstrated excellent consistency between full-dose and 10% dose images (mean DSC 0.997±0.002, HD95 0.73±0.26 mm). In contrast, 5-lobe segmentation showed substantially lower accuracy and greater variability, even at routine dose. Automated tissue classification revealed a mean decrease in functional (inflated) lung volume of 17.12%±5.69% in the right lung and 17.15%±5.56% in the left lung, accompanied by marked increases in emphysema, infiltrated, and collapsed volumes. Conclusion: AI-based whole-lung segmentation remains highly accurate at 10% of the routine dose, supporting its potential for lung transplant volumetry. However, tissue classification and 5-lobe segmentation exhibited notable degradation, highlighting limitations of existing tools at ultra-low dose levels.

Background: Evaluation of middle ear prostheses requires high-resolution imaging to assess position and structural integrity. Purpose: To assess the diagnostic performance of photon-counting-detector (PCD)-CT for middle ear prosthesis imaging in cadaveric specimens. Materials and Methods: Nine different middle ear prostheses were implanted in 5 fresh-frozen whole-head specimens, both in correct anatomic position and displaced to mimic malposition. Specimens were scanned on PCD-CT (NAEOTOM Alpha, Siemens) and energy-integrating-detector (EID)-CT (SOMATOM Force, Siemens), reconstructed at 0.2 mm and 0.4 mm, respectively. Eighteen cases (9 prostheses, 2 positions) per scanner were evaluated by 2 neuroradiologists (36 readings), blinded to scanner and position, who scored prosthesis position, reader confidence, and overall image quality. Agreement with the reference standard and between readers was compared. Results: PCD-CT showed higher diagnostic confidence compared with EID-CT, while accuracy and inter-reader agreement did not demonstrate significant differences between scanners. Correct readings, PCD-CT versus EID-CT: lateral contact, 31/36 (86.1%; 95% CI, 58.3-94.4) vs 29/36 (80.6%; 95% CI, 55.6-91.7), P = .500; medial contact, 30/36 (83.3%; 95% CI, 55.6-94.4) vs 26/36 (72.2%; 95% CI, 50.0-83.3), P = .250. PCD-CT identified all cases of complete contact loss (score 3), while EID-CT missed 1 reading per contact. Inter-reader agreement may be higher with PCD-CT vs EID-CT (lateral: κw: 0.804 [95% CI, 0.578-0.941] vs 0.571 [95% CI, 0.302-0.778], P = .250; medial: 0.684 [95% CI, 0.191-0.934] vs 0.481 [95% CI, 0.024-0.835], P = .250). Reader confidence was significantly higher with PCD-CT; mean paired differences for Readers 1 and 2 were 9.4 (95% CI, 1.8-17.1; Holm-adjusted P = .019) and 23.1 points (95% CI, 13.4-32.7; P < .001) for lateral contact, and 16.4 (95% CI, 10.0-22.8; P < .001) and 17.8 points (95% CI, 6.8-28.8; P = .007) for medial contact. Readers preferred PCD-CT image quality in 16/18 (88.9%; 95% CI, 65.3-98.6) and 17/18 cases (94.4%; 95% CI, 72.7-99.9), respectively. Conclusion: PCD-CT was associated with significantly higher reader confidence and preferred image quality, without significant differences in diagnostic accuracy or inter-reader agreement. These findings suggest that PCD-CT is a promising modality for postoperative middle ear evaluations and warrants further investigations.

Concerns about the risk of radiation from CT have driven a spectrum of major advances in radiation dose reduction technology since the 2000s, including added beam filtration, dynamic z-axis collimation, automatic tube current modulation and tube potential selection, advanced iterative reconstruction, and deep learning-based reconstructions. The introduction of photon-counting detector CT further improved dose efficiency through electronic noise rejection, optimal photon-energy weighting, and high-resolution data acquisition. This Review summarizes key milestones and technologic innovations that have led to substantial reductions in CT radiation doses and provides examples of the cumulative impact of these advances on CT doses and image quality for routine examinations in various anatomic sites. This article also highlights the importance of an objective task-based image quality assessment to ensure that diagnostic performance is not compromised as dose is reduced, particularly for emerging deep-learning-based reconstruction and postprocessing noise reduction methods.

Objective: To develop an architecture-agnostic framework that estimates, calibrates, and leverages total uncertainty (aleatoric + epistemic) in pre-trained, deep-learning denoising models for low-dose computed tomography (CT). Methods: Aleatoric and epistemic uncertainties were estimated using physics-based inference-time augmentation and training-free, post-hoc Monte Carlo dropout, respectively, followed by non-parametric re-calibration for improved uncertainty calibration. To leverage uncertainty, we explored adaptive local fusion (ALF) guided by local mean-to-uncertainty ratio. For proof-of-concept, this framework was assessed using pre-trained U-net and ResNet-based models across datasets varying in CT tasks, radiation dose, and lesion characteristics. Uncertainty estimation and calibration were assessed in cadaver scans, using normalized-root-mean-square-error (NRMSE) and normalized-calibration-error (NCE), respectively. ALF was evaluated with chest and liver exams, using noise, structural similarity index (SSIM), and lesion detectability. Lesion detectability was quantified using clinically validated deep-learning model observer, with Wilcoxon signed-rank test to assess significance. Results: This framework provided accurate uncertainty quantification and calibration: NRMSE range [1.2%, 2.4%], NCE [0.9%, 2.2%]. Compared to original pre-trained models, ALF yielded comparable or lower noise, improved lesion structural fidelity and detectability (p<0.05): For lung nodules - noise reduction up to 69.7%, SSIM range (ALF vs pre-trained) [0.92, 0.96] vs [0.78, 0.87], detectability improvement up to 12.9%; for liver metastases - noise reduction up to 35.0%, SSIM range (ALF vs pre-trained) [0.82, 0.99] vs [0.80, 0.98], detectability improvement up to 13.2%. Conclusion: Our framework effectively benchmarked and utilized total uncertainty to enhance diagnostic image quality with pre-trained CT denoising models. Significance: This framework can facilitate performance monitoring, deployment optimization, and trustworthiness establishment

Objective: This study assessed CT number consistency in different materials and tissues as x-ray tube potential, scanner model and manufacturer, and phantom size were varied. We investigated using virtual monoenergetic imaging (VMIs) to improve CT number consistency compared to single-energy (SE) images. Approach: Four clinically-relevant material inserts (calcium, adipose, iodine, soft tissue) were placed into two phantom sizes (28-cm and 40-cm width) and scanned at each available tube potential on three CT scanners with distinct methods of spectral data collection. SE images and VMIs were reconstructed for each scan. The average CT number of each insert was calculated for each image series, and the consistency of this average value across different combinations of scan parameters was compared between VMIs and SE images using coefficient of variation (COV), median, range and a linear model estimating scanner-based CT number differences. Main results: The COV for calcium, adipose, iodine and soft tissue inserts decreased from 10.1%, 9.7%, 28.8%, and 6.2% in SE images to 1.6%, 3.1%, 2.9%, and 5.7% in VMIs, respectively. When only tube potential was changed, the median standard deviation in insert average CT number decreased in VMIs compared to SE images in all materials. VMIs decreased the range of insert values between phantom sizes. VMIs decreased CT number variation due to scanner in both phantom sizes for iodine and in the calcium, 28-cm phantom. Inter-manufacturer differences in VMI CT numbers were observed in low density materials and the 40-cm phantom. Significance: Scan parameter variation affects CT numbers, negatively impacting quantitative imaging consistency and downstream patient metrics. This limits measurement standardization in serial studies, diverse cohorts, and between hospitals with data collection differences. VMIs from spectral CT can reduce this variability by providing more consistent material CT numbers relative to changes in tube potential, scanner model, and patient size.

Photon-counting detector CT (PCD-CT) directly detects individual photons and measures their energies, enabling the acquisition of multi-energy data from a single x-ray source with higher spatial resolution, thereby allowing visualization of smaller arteries and veins and other important imaging features not seen at conventional CT. It also substantially reduces calcium and stent blooming artifacts. Appropriate patient selection and task-based optimization of acquisition, reconstruction, and interpretation are critical to realize these benefits. Vascular imaging examinations for which early evidence shows that PCD-CT demonstrates higher diagnostic performance compared to conventional CT include coronary CT angiography (CTA) for evaluation of partially calcified plaques or stents, peripheral runoff CTA, detection of cerebrospinal fluid-venous fistulas, and head CTA for evaluation of small intracranial vessels and aneurysms. Current limitations include excessive image noise with the sharpest reconstruction kernels and challenges when adapting conventional CT protocols (eg, requiring sharper kernels, thinner slices, higher matrices, tailored scan modes) to achieve time-efficient image reconstruction and hanging protocols. New solutions and directions will include deep-learning-based image noise reduction, improved spectral separation, strategic triage of specific patient populations to PCD-CT systems, and subspecialty clinical practice recommendations for this new technology.

A deep learning–based neural network enabled virtual monoenergetic imaging at energy-integrating detector CT, reduced blooming artifacts, and improved stenosis assessment in heavily calcified coronary plaques, with performance comparable to ultrahigh-resolution photon-counting detector CT.