Application of deep-learning-based super-resolution image quality improvement on vessel wall imaging

Authors

    Seokhee Park1, Minkook Seo1, Woojin Jung2, Kook-Jin Ahn1, Bum-soo Kim1, Jinhee Jang1


     

    1The Catholic University of Korea, Seoul St. Mary’s Hospital, Korea, Republic of
    2AIRS Medical Inc, Korea, Republic of
    [email protected]

    KCR (2023)

    PURPOSE

    Obtaining high resolution and acceptable signal-to-noise ratio (SNR) in intracranial vessel wall imaging (VWI) is a challenging task. We applied a deep-learning (DL)-based method on VWI which enables both super-resolution (SR) and noise reduction and explored the clinical potential of this model.

    MATERIALS AND METHOD

    117 patients with VWI were included, scanned by a 3T MRI (MAGNETOM Vida, Siemens Healthineers) with 3D TSE T1-weighted images (T1WI). Acquisition parameters were; TR/TE = 800/23ms, ETL 40, black blood preparation, acceleration factor 4, voxel 0.51×0.51×0.45mm3, FOV 180x180x80mm3, scan time 5min 52sec.
    An improved DL-based algorithm which simultaneously applies SR and denoising was trained using about 3500 MRIs acquired by various 3D sequences (SwiftMR™, AIRS Medical). The DL model was based on U-net architecture, which receives original VWI (Conv-VWI) and produces DL-augmented super-resolution VWI (SR-VWI) with voxel size of 0.28×0.28×0.45mm3.
    We assessed both image quality and vessel wall lesions by comparing Conv-VWI and SR-VWI. Quality rating was done for overall image quality, visual SNR, vessel conspicuity, artifact; for wall lesions, confidence of detection, lesion conspicuity, and contrast enhancement were evaluated. Normalized signal intensities (SI) and standard deviation (SD) of the normal vessel walls and lumens of MCA M1 segment (M1) and mid-basilar arteries (BA) were measured, with calculation of SNR and contrast-to-noise ratio (CNR). Same procedure was done to wall lesions on both T1WI and contrast-enhanced T1WI (CE-T1WI).

    RESULTS

    Scores for overall image quality (4.2±0.7 vs. 2.9±0.6), visual SNR (4.3±0.7 vs. 2.6±0.6), vessel conspicuity (4.3±0.7 vs. 3.1±0.7), confidence of lesion detection (4.6±0.6 vs. 2.6±0.8) and lesion conspicuity (4.3±0.7 vs. 3.0±0.8) were all significantly higher in SR-VWI than Conv-VWI (P<0.01). SNR (M1 1.2 times, BA 1.6 times) and CNR (M1 1.1 times, BA 1.4 times) of normal walls were significantly higher in SR-VWI than Conv-VWI (P<0.01). Same results were observed in lesion assessment (SNR 1.6 times, CNR 1.5 times; CE-T1WI SNR 2.7 times, CNR 2.6 times, P<0.01). Degree of lesion enhancement was slightly more prominent on SR-VWI compared to Conv-VWI (visual scores 1.4±0.7 vs. 1.2±0.8, normalized SI on CE-T1WI 103.5±40.5 vs. 100.1±36.5).

    CONCLUSION

    DL-based model with super-resolution and denoising provides better image quality, higher SNR and CNR with more confident detection of the vessel wall lesions and more noticeable contrast enhancement.