292 - Clinical Impact of an AI-based Artifact Reduction Reconstruction for Cone Beam CT Angiography During Hepatic Radioembolization

Wednesday, April 15, 2026

12:18 PM - 12:27 PM ET

Author/Co-author(s)

PB

Paul Beunon, MD

Faculty, Surgical Radiology
Gustave Roussy, Villejuif, France, France

Presenting Author(s)

Adam D. Talenfeld, MD, MS, FSIR

Associate Professor of Clinical Radiology
Weill Cornell Medicine, United States

Author/Co-author(s)

RG

Ruben Geevarghese, MD

Senior Research Scientist
Memorial Sloan Kettering Cancer Center, United States
Mahmoud Odeh, MD

Postdoctoral research fellow
Beth Israel Deaconess Medical Center, United States
Brett Marinelli, MD

Assistant Attending
Memorial Sloan Kettering Cancer Center, United States
AO

Alfredo Ormeno Zuniga, BS

Research coordinator
Program for Intelligent Image-guided Interventions NewYork-Presbyterian/Weill Cornell Medical Center, United States
LT

Lambros Tselikas, MD

MD
Gustave Roussy Cancer center, France
FD

Frederic Deschamps, MD

MD
Gustave Roussy, France
MR

Maxime Ronot, MD
Kyungmouk Steve Lee, MD

Department of Radiology, Division of Interventional Radiology, Assistant Professor
Weill Cornell Medicine, United States
MA

Muneeb Ahmed, MD, FSIR

Professor of Radiology
Beth Israel Deaconess Medical Center/Harvard, United States
Ammar Sarwar, MD, FSIR (he/him/his)

Associate Professor of Radiology
Beth Israel Deaconess Medical Center, United States
MD

Marco Dioguardi Burgio, MD

Faculty, Division of Interventional Radiology
Hopital Beaujon, France

Purpose: To evaluate the impact of a new AI-based reconstruction algorithm in reducing streak artifacts related to arterial pulsatility and contrast flow variability in contrast-enhanced cone beam CT angiography (CBCT) during mapping for liver radioembolization.

Materials and Methods: From the international prospective multicenter PALAS-VP registry {1}, CBCTs were acquired from July, 2024 to July, 2025 with IGS angiography systems and reconstructed on Advantage workstations both with and without CleaReconDL AI-based software (all hardware and software GE HealthCare, Chicago, IL). The first 10 consecutive patients from each of 5 centers in the United States and France were included, regardless of CBCT quality. Proximal (whole liver or lobar) and selective CBCTs (injecting at the site of planned radioembolization) were performed in each patient, yielding a total of 100 CBCT datasets. Three senior interventional radiologists blinded to reconstruction type independently assessed each CBCT using 5-point Likert scales for target artery delineation, non-target artery depiction, lesion visualization, treatment zone parenchymal delineation (selective CBCTs only), and overall image quality. Information loss was also evaluated, as was the presence of any AI hallucinations. Scores were averaged across readers and compared using Wilcoxon signed-rank tests. Results were evaluated across the five centers to assess inter-center variability.

Results: In both proximal and selective CBCTs, AI-based reconstruction significantly improved all assessed parameters. In proximal CBCTs, AI-based reconstruction improved target and non-target arterial delineation (3.4 to 4.4 and 3.3 to 4.4), lesion visualization (3.1 to 3.7), and overall image quality (3.2 to 4.2), all p< 0.001. In selective CBCTs, CleaReconDL improved target and non-target arterial depiction (3.2 to 4.2 and 3.4 to 4.2), lesion visualization (3.2 to 3.7), treatment zone parenchymal delineation (3.4 to 3.8), and overall image quality (3.0 to 3.8), all p< 0.001. No clinically relevant anatomical or lesional information was lost with the AI-based reconstruction, nor were any AI hallucinations noted. There was no significant variability in outcomes between centers.

Conclusion: AI-driven reconstruction significantly enhances both proximal and selective CBCT image quality and visualization of key features relevant to hepatic radioembolization without loss of any diagnostic information or creation of new artifacts.