Click here to

Session: Imaging General ePoster Viewing [Return to Session]

3D Super-Resolution Reconstruction for Hippocampus Sparing in Brain Radiotherapy (in Silico Study)

Juyoung Lee1,2*, Jaehee Chun2,3, Jin Sung Kim2,3, Taeho Kim1, (1) Washington University School of Medicine in St. Louis, St. Louis, MO, (2) Yonsei University College of Medicine, Seoul, KR, (3) Oncosoft Inc., Seoul, KR

Presentations

PO-GePV-I-41 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Contrast-enhanced MRI is a standard procedure in brain radiotherapy. Soft-tissue contrast enhancement is beneficial to tumor delineation, while hippocampus contouring is challenging. Non-contrast enhanced MRI, however, is useful for hippocampus contouring, often omitted from the standard protocol due to scan time limitation and patient compliance. In this silico study, we tried to meet this challenge by securing high-resolution (HR) non-contrast images from low-resolution (LR) images by using interpolation and deep learning-based super-resolution network.

Methods: Fifteen scans of HR brain MRIs (256x256x200) from 1.5T Philips Ingenia MRI were prepared as a training dataset after paired with down-sampled images by 1/4. Artificial HR non-contrast images from the LR images (64x64x50) were yielded by interpolation (64x64x200) and deep super-resolution network (256x256x200), consequently. Three indexes to quantify the spatial quality (BRISQUE), naturalness (NIQE), and perceptioness (PIQE) as well as the structural similarity index measure (SSIM) were used to evaluate the results.

Results: Our model reconstructed the non-contrast MR images in 3D shape with four times higher resolution on three patients, two different sequences of images. For each result, the images were compared globally and locally close to the hippocampi. Mean±Std of SSIM on each plane was reported: 0.823±0.078 (axial), 0.842±0.068 (sagittal), 0.825±0.068 (coronal) from the proposed results, while 0.773±0.084, 0.783±0.087, 0.745±0.075 from the LR images (overall increase of 9.5%). In the no-reference indexes, each result showed a value closer to the reference HR images. For each plane, 1) BRISQUE: 45.21±6.55→41.91±5.51 (HR: 40.89±5.92), 42.85±8.08→41.28±5.12 (HR: 31.22±9.91), 46.94±6.55→40.27±5.72 (HR: 37.11±8.14). 2) NIQE: 18.87±0.00→5.15±1.36 (HR: 5.64±1.72), 18.88±0.00→6.22±0.70 (HR: 4.44±1.02), 18.87±0.00→6.45±1.32 (HR: 4.97±1.54). 3) PIQE: 66.99±10.86→29.34±4.49 (HR: 31.86±4.11), 68.82±8.59→50.30±8.00 (HR: 31.79±4.05), 73.56±11.38→75.72±9.11 (HR: 35.73±6.78) was recorded.

Conclusion: Through this study, HR non-contrast MRIs could be generated without lengthy scans, and made it possible to easily distinguish the hippocampus in brain radiation therapy.

Keywords

3D, MRI, Reconstruction

Taxonomy

IM- MRI : Image Reconstruction

Contact Email

Share: