In a retrospective, comparative, and blinded multicenter reader study published in CVIR Endovascular, an artificial intelligence-based image processing platform designed for two-dimensional angiography achieved non-inferior overall image quality in 98.9 percent of clinical evaluations compared with its predecessor system. The investigation, led by Dr. David C. Madoff alongside an international team of clinical specialists, evaluated the OPTIQ AI spatial denoising algorithm against the previous-generation OPTIQ platform across five separate imaging modes and seven distinct anatomic territories. Thirteen board-certified physicians conducted independent, paired assessments of 553 patient image series, determining that the deep learning algorithm matched or exceeded the baseline platform across routine endovascular workflows while showing demonstrable gains in image contrast, noise suppression, and diagnostic confidence in lower-dose operational modes.
Endovascular interventions have expanded steadily in both overall procedural volume and anatomical complexity, propelled by documented clinical benefits that include reduced procedural trauma, lower complication rates, and shorter hospital recovery periods. These minimally invasive procedures are routinely performed on elderly individuals presenting with extensive vascular comorbidities, as well as on younger pediatric patients who may require repeated, staged catheter-based therapies across their lifetimes. Because these interventions depend fundamentally on continuous X-ray visualization for catheter manipulation and therapeutic deployment, their increasing adoption has contributed to cumulative occupational and patient radiation exposure across cardiology, vascular surgery, interventional radiology, and neurointerventional specialties. As a consequence, adherence to the As Low As Reasonably Achievable principle remains a central imperative, requiring engineering solutions that optimize image clarity without escalating radiation exposure.
The physical basis of angiographic image optimization has historically centered on refinements to X-ray tubes, generator systems, and flat-panel digital detectors. Recent progress has expanded beyond hardware boundaries by integrating contrast-to-noise ratio-driven exposure control and material-specific beam filtration, which dynamically adapt tube voltage, pulse duration, electrical current, focal spot geometry, and filtration to anatomical targets. Image processing represents a complementary operational tier, particularly through spatial denoising algorithms that suppress electronic and stochastic noise within individual image frames while preserving fine anatomical margins, vascular borders, and radiopaque device details. The OPTIQ AI platform replaces conventional analytical filtration with a deep convolutional architecture trained on extensive image datasets across diverse radiation doses and beam spectra, learning detector responses, spatial pixel correlations, and noise properties independent of specific clinical structures.
To evaluate whether these machine learning architectures translate into verifiable improvements under clinical conditions, the investigators executed a retrospective, multicenter evaluation protocol operating under an institutional review board exemption waiver from Western Copernicus Group. The study gathered anonymized clinical imaging records from eleven participating academic and clinical medical centers situated across the United States and the European Union, complying fully with the Health Insurance Portability and Accountability Act and General Data Protection Regulation mandates. Clinical data management was maintained through an electronic data capture platform governed by strict administrative access protocols, ensuring that data collectors, image processors, reading coordinators, and trial biostatisticians remained segregated to prevent premature unblinding throughout case logging, automated reprocessing, randomization, and subsequent score de-randomization.
The study cohort comprised 182 distinct patient subjects, including 93 male individuals, 80 female individuals, and 9 patients whose demographic sex was unrecorded in the registry. Within this population, 29 subjects representing 16 percent of the cohort were younger than 18 years of age, while 153 subjects representing 84 percent were adults. The assembled imaging repository included 553 total image series, of which 345 series originated from clinical sites in the United States and 208 series were acquired across European institutions. In total, 51 distinct procedural types were captured across seven defined anatomical compartments: the head, neck, adult heart, pediatric heart, thorax, abdomen, and extremities, providing a heterogeneous foundation representing interventional radiology, neurointerventions, adult cardiology, and pediatric structural heart procedures.
Local clinical specialists identified at least one representative, clinically meaningful recording per procedural acquisition mode, incorporating orthogonal projections for biplane roadmapping examinations. The selection protocol was deliberately neither consecutive nor randomized, and local physicians operated without instructions to bias selections toward unusually complex anatomy. Selected scenes were categorized by clinical difficulty into difficult or non-difficult imaging conditions, and visible endovascular hardware was documented and stratified into high or low X-ray visibility classes. Selected raw projection data underwent automated reprocessing using manufacturer-default clinical protocols for both the OPTIQ AI architecture and the predecessor OPTIQ platform, after which scene pairs were assigned randomized identifiers to obscure the underlying processing pipeline during side-by-side diagnostic evaluations.
Reader assessments were conducted by thirteen board-certified medical specialists possessing a median clinical practice experience of 19 years, spanning a range of 7 to 27 years. The reading panel included three interventional radiologists assigned to abdominal and thoracic series, three neurointerventionalists assessing cranial and cervical vascular runs, and seven cardiovascular specialists evaluating peripheral, adult cardiac, and pediatric cardiac examinations. Four of the participating physician readers had also assisted during the preliminary institutional data collection phase. Every reader underwent standardized pre-study orientation using separate illustrative training images excluded from the formal testing dataset, viewing synchronized image pairs side by side on calibrated diagnostic workstations with randomized left-right monitor orientation.
The readers evaluated each paired scene using standardized five-point Likert scales covering overall image quality, contrast of clinically relevant anatomical details, noise reduction efficiency, and physician confidence in assessing delicate vessel and device margins. For overlay reference sequences, contrast of the static vascular roadmap and contrast of active catheter devices were graded independently to assess the specialized overlay algorithm. The primary study endpoint was defined as non-inferiority in overall image quality, requiring the lower bound of the 95 percent confidence interval for combined ratings of identical, superior, or much superior quality to reach or exceed 70 percent, while restricting ratings of much lower quality to less than 5 percent across the overall cohort.
In the primary endpoint evaluation, OPTIQ AI satisfied both non-inferiority conditions, achieving equivalent or superior overall image quality in 98.9 percent of all 1,659 paired reader evaluations, with a 95 percent confidence interval spanning 98 percent to 100 percent. The proportion of readings judging the AI system as much lower than the predecessor was limited to 0.5 percent, with an estimated confidence interval between 0.01 percent and 0.8 percent. Across all paired scoring interactions, readers classified the AI algorithm as much higher in 34.1 percent of instances, higher in 49.1 percent, equivalent in 15.7 percent, lower in 0.6 percent, and much lower in 0.5 percent. Subpopulation analyses confirmed consistent non-inferiority across patient body mass index, ethnic background, chronological age, geographic site origin, technical case difficulty, device visibility, and observer experience levels, with subgroup non-inferiority rates ranging between 96.5 percent and 100 percent.
Secondary endpoints were evaluated through a predefined hierarchical gate-keeping strategy to test statistical superiority across individual metrics and acquisition modalities. Statistical significance required rejecting null hypotheses stating that the combined probability of higher or much higher Likert ratings was 50 percent or lower, using generalized estimating equations to account for repeated-measures reader correlation. The first 22 hierarchical endpoints successfully demonstrated statistical superiority for OPTIQ AI across fluoroscopy, native acquisition, roadmap, and overlay reference modes. Furthermore, the specialized extended overlay reference algorithm satisfied its secondary objective, concurrently improving vessel map contrast while preserving fine device visibility in live fluoroscopic guidance.
The formal hierarchical testing sequence ceased when overall image quality in digital subtraction angiography was examined. Although 52.3 percent of digital subtraction evaluations favored the AI platform compared to only 0.7 percent favoring the predecessor, the lower bound of the corresponding 95 percent confidence interval reached 46.6 percent, failing to clear the mandatory 50 percent threshold for statistical superiority. Consequently, the remaining three secondary hypotheses in digital subtraction angiography were not formally tested under the gate-keeping protocol. The authors attributed this outcome to the high baseline signal-to-noise ratio intrinsic to high-dose subtraction imaging, which produces a subjective perceptual ceiling effect that limits noticeable cosmetic improvements compared to low-dose fluoroscopic sequences.
Pairwise inter-reader agreement analysis demonstrated complete scoring concordance in 49.3 percent of paired evaluations, while 45.4 percent of paired scores differed by exactly one Likert scale category, meaning that 94.7 percent of reader assessments diverged by no more than a single rating division. Significant rating divergence was infrequent, with two-category differences occurring in 4.2 percent of evaluations, three-category differences in 0.5 percent, and four-category differences in 0.6 percent. While reader agreement rates showed no statistically significant disparity across imaging modalities, concordance differed significantly across anatomical zones, with readers showing highest consensus in cervical neck examinations at 61.1 percent and lowest agreement in extremity and thoracic acquisitions at 43.7 percent and 42.5 percent, respectively.
The authors outlined several methodological constraints inherent to the study design, including its reliance on qualitative visual impressions rather than standardized physical metrics such as contrast-to-noise ratios, spatial resolution measurements, or objective noise power spectra. Additionally, four of the thirteen physician readers participated in preliminary clinical data gathering, case selection was directed by local physician judgment rather than strict consecutive enrollment, and identical raw detector exposures were utilized to reconstruct both processing conditions, precluding direct measurement of procedural radiation dose reductions. While earlier benchtop and animal studies demonstrated radiation dose reductions of up to 83 percent using contrast-to-noise exposure controls, prospective clinical trials will be necessary to verify whether the visual enhancements delivered by deep learning spatial denoising permit operational dose reductions during active interventions.
Subject of Research: Clinical evaluation of deep learning-based image processing in two-dimensional angiography
Article Title: Clinical evaluation of a novel artificial intelligence-based image processing algorithm for two-dimensional angiography
Article References: Madoff, D. C., Aguilar, H. T., Angtuaco, M. J., Asnes, J. D., Barr, J. D., Frei, D. F., Jr., George, J. C., Hohenwalter, E. J., Kaestner, T., Lacey, D. L., Meraj, P., Obuchowski, N., Puma, J. A., Quinn, B. P., Sahlein, D. H., Secemsky, E. A., Setser, R. M., & Tashenov, S. (2026). Clinical evaluation of a novel artificial intelligence-based image processing algorithm for two-dimensional angiography. CVIR Endovascular, 9(1), Article 145. https://doi.org/10.1186/s42155-026-00785-6
Image Credits: AI Generated
DOI: 10.1186/s42155-026-00785-6
Keywords: angiography, image processing, artificial intelligence, deep learning, fluoroscopy, digital subtraction angiography, radiation protection, interventional radiology, cardiology, neurointervention, spatial denoising, image quality, multicenter study, reader evaluation, OPTIQ AI, ALARA principle
News Source: Blake Davidson. (November 18, 2026). Deep Learning Image Processing Matches Predecessor Quality Across Diverse Angiographic Procedures. Scienmag.



