Comparison of Models in Deep Learning-Based Synthetic Digital Subtraction Angiography Generation

Monday, March 31, 2025

3:18 PM – 3:27 PM CT

Location: Music City Center, 205 A

Presenting Author(s)

JH

Joshua Halpern

Researcher
MedStar Georgetown University Hospital, United States

Author/Co-author(s)

MB

Muhammad Bilal, PhD

Professor of Applied AI and Technology Ethics
Birmingham City University, United Kingdom
RF

Ross Filice, MD

Director of Research
MedStar Georgetown University Hospital, United States
Nariman Nezami, MD (he/him/his)

Associate Professor
Georgetown University Medical Center, United States
GS

Gajan Sivananthan, MD

Assistant Professor of Radiology
MedStar Washington Hospital Center, United States
Emil I. Cohen, MD, FSIR

Associate Professor
Medstar Georgetown University Hospital, United States

Purpose: To develop and evaluate a U-Net based deep learning model to generate 2-dimensional digital subtraction angiography images from native angiograms and compare it with a previously published GAN model.

Materials and Methods: Two-dimensional digital subtraction angiography (2D-DSA) images and their corresponding native angiography images were collected from patients undergoing angiographic procedures between July 1, 2023, and October 1, 2023. A total of 44 image sets (native and subtraction) from 44 patients (19 male, 25 female) were obtained. The dataset was augmented by applying magnification and rotation to the original images, resulting in 3,096 total images. These were randomly split into training and validation sets using a 9:1 ratio. An additional eight images from new patients were collected for the test set. Two deep learning models—a generative adversarial network (GAN), previously described in the literature, and an encoder-decoder convolutional neural network (U-Net)—were trained on the collected data to process native angiographic images, removing background structures to generate synthetic digital subtraction angiography (SDSA) images. The resulting images from each model were qualitatively assessed by 11 interventional radiologists, who made binary selections to determine the clinically preferable image blindly.

Results: U-Net output images showed statistically significant improvements in Deep Feature Similarity (0.67 ± 0.09 vs. 0.47 ± 0.05) , pHash Similarity (0.78 ± 0.09 vs. 0.71 ± 0.09), and HOG Similarity (0.86 ± 0.02 vs. 0.80 ± 0.01) over GAN output images. Significant differences in SSIM and PSNR were also noted. Qualitative visual comparison of each model’s outputs by interventional radiologists showed a strong preference for U-Net SDSA images.

Conclusion: The developed U-Net model generated higher quality synthetic digital subtraction angiography images than the GAN model given the same training data. Future work will explore clinical validation of the U-Net model to determine its impact on diagnostic accuracy and procedural planning in real-world settings, with an emphasis on radiation dose reduction and clinical decision-making.