Image quality improvement for digital subtraction angiography (DSA) around high-density obstacles using deep-learning: A bench-top study.

Sunday, March 5, 2023

4:21 PM – 4:30 PM

Location: Phoenix Convention Center, 224AB

Presenting Author(s)

Kentaro Yamada, MD, PhD (he/him/his)

Research Assistant Professor
Oregon Health & Science University

Disclosure(s): No financial relationships to disclose

Author/Co-author(s)

TT

Takeshi Takata, PhD

Assistant Professor
Teikyo University
AA

Arianna Anoushiravani, BS

Rsearch fellow
Oregon Health & Science University
LC

Leonardo Campos, MD

Interventional Radiologist
Oregon Health & Science University
TG

Todd Graham, MS, BS

Senior Research Associate
Oregon Health & Science University
JK

Jun'ichi Kotoku, PhD

Professor
Teikyo University
Khashayar Farsad, MD, PhD, FSIR (he/him/his)

Professor
Oregon Health and Science University

Disclosure(s):

Kentaro Yamada, MD, PhD: No financial relationships to disclose

Khashayar Farsad, MD, PhD, FSIR: Auxetics, Inc.: Ownership Interest (Ongoing); Baylis Medical Technologies: Consultant (Ongoing); Cook Medical: Consultant (Ongoing); Eisai Co., Ltd: Advisory Committee or Review Panel Member (Terminated, April 15, 2022); Guerbet LLC: Research Grant Recipient (Ongoing); Inquis Medical: Advisory Committee or Review Panel Member (Terminated, May 1, 2022); J&J / Ethicon / NeuWave: Consultant (Ongoing), Speaking and Teaching (Ongoing); OEC: Contracted Research (Ongoing); RenovoRx: clinical trial site (Ongoing); W. L. Gore: Research Grant Recipient (Ongoing)

Purpose: DSA image quality is often degraded beneath high-density obstacles such as a contrast-filled bladder, renal pelvis, or metal implants. Deep-learning (DL) has been utilized to improve image quality with noise reduction and increased resolution in CT and MRI; however, DL for image optimization in angiography is not prevalent. This study aims to evaluate the efficacy of DL-based algorithms to improve DSA images impaired with overlying high-density obstacles using a bench-top model.

Materials and Methods: A bench-top circulation model was made with a cerebral angiography phantom and a circulation pump. Standard DSA as ground truth and DSA impaired with an overlying syringe filled with 20% iodinated contrast were taken as paired images. A total of 23 DSA series of 2715 paired images were used to train the DL model, while 5 DSA series of 587 images were taken as validation data sets. Validation data sets were processed via the trained DL-based model, and DL-based subtraction angiography (DLSA) data was generated. Multiply-processed DSA (MP-DSA) data were also generated by multiplying each pixel value of the impaired image and DLSA utilizing DLSA as an image filter. DLSA and MP-DSA images were quantitatively evaluated by structural similarity index measure (SSIM). Images were compared to the ground truth for quality evaluation of vascular visualization and clinical usefulness by three interventional radiologists with a 5-point numerical scale (1, poor; 5, excellent) {1}. SSIM of DLSA and MP-DSA were compared to that of the impaired DSA using Wilcoxon signed-rank test.

Results: Median SSIM (IQR, p-value vs impaired DSA, Cohen's d vs impaired DSA) of impaired DSA, DLSA and MP-DSA compared to ground truth are 0.95 (0.94–0.95), 0.96 (0.94–0.95, p < 0.01, Cohen's d = 1.16), 0.89 (0.88–0.89, p < 0.01, Cohen's d = 9.19). Median (IQR) vascular visualization and clinical usefulness scores were as follows: DLSA; 4 (3–5), 4 (3.5–4.5): MP-DSA; 4 (3.5–5), 4 (3.5–5).

Conclusion: The DL model provided improved image quality for degraded DSA images behind a high-density obstacle. SSIM values were lower with MP-DSA, likely due to contrast resolution; however, MP-DSA and DLSA both scored higher on visual evaluation scores.