ESTRO 2024 - Abstract Book

S1379

Clinical - Head & neck

ESTRO 2024

1 Hong Kong Sanatorium and Hospital, Research, Hong Kong, Hong Kong. 2 The Chinese University of Hong Kong, Imaging and Interventional Radiology, Hong Kong, Hong Kong. 3 Hong Kong Sanatorium and Hospital, Diagnostics and Interventional Radiology, Hong Kong, Hong Kong

Purpose/Objective:

The use of MR simulators in radiotherapy has recently gained immense popularity owing to the superior soft-tissue contrast of MRI as compared to CT scans [1]. Unlike diagnostic MRI, MR images from MR simulators are mainly used for delineating targets and organs at risk (OARs) instead of diagnosing diseases. Therefore, high-resolution images with a large field of view are commonly used. However, acquiring images with high resolution comes at the cost of lengthening its image acquisition time, which can lead to intra-session motion and increase patient discomfort along with operational costs. Hence, there is a need for MR sequences that offer high resolution, low geometric distortion, broad coverage, and short scanning time. With the advancement of technology, especially in the application of deep learning in the medical field, deep learning-based super-resolution post-processing algorithms have been developed. These algorithms could generate high-resolution images from lower-resolution images with the addition of reducing the image noise, which could improve the image quality and signal-to-noise ratio in MR images without increasing scan time [2]. However, the clinical value of these deep learning-based super-resolution algorithms for radiotherapy purposes has not been investigated. Therefore, this study aims to evaluate the potential value of deep learning-based super-resolution generated T1W 2D turbo-spin-echo (TSE) images for radiotherapy by comparing the inter-observer delineation repeatability to the conventional 3D-T1W gradient echo images. Twelve patients (8 males, 51±11 years; 4 females, 51±19 years), referred by clinicians for either post-radiotherapy follow-up or with suspicious tumors in the head and neck region, were retrospectively included. They underwent clinical 1.5T MR scans in the head-and-neck region. As part of the clinical MRI protocol, a post-contrast deep learning-based TSE scan (2D-TSE, axial, TR/TE = 570/9.2 ms, reconstruction voxel size = 0.46 x 0.46 x 4 mm³, Fat-Sat = ON, acquisition time = 113s) and a post-contrast conventional gradient echo scan (3D-VIBE, axial, TR/TE1/TE2 = 7.1/2.37/4.77 ms, reconstruction voxel size = 0.74 x 0.74 x 2 mm³, DIXON = ON, acquisition time = 180s) were acquired using a 20-channel head and neck coil. Delineations of various OARs were carefully conducted by two experienced MR physicists, each with over 7 years of experience. The OARs included the lymph node, left and right parotid gland, pituitary gland, and brainstem. To ensure consistency, a common consensus was reached for each OAR delineation, resulting in 2 segmentations per OAR per image. The inter-observer delineation similarity was assessed using the dice-similarity coefficient (DSC), volume difference, and Hausdorff distance. Kruskal Wallis tests, and Wilcoxon signed-rank test were conducted to compare the inter-observer delineation similarity parameters between the deep-learning-based 2D-TSE scan (DL) and the conventional 3D-VIBE scan (conventional) among different OARs delineations. The significance level was set at 0.05. Material/Methods:

Results: