Artificial Intelligence-Assisted Compressed Sensing Technique Accelerates Magnetic Resonance Imaging Simulation for Head and Neck Cancer Radiation Therapy

Abstract

Purpose

To explore the potential of artificial intelligence-assisted compressed sensing (ACS) technique, when compared with that of conventional parallel imaging (PI) technique, in magnetic resonance imaging (MRI) simulation for head and neck cancer radiation therapy.

Methods and Materials

Fifty-two patients with pathologically confirmed head and neck cancer underwent MRI simulation using a 3.0-T MRI simulation system. For each patient, axial T1-weighted gradient spin echo, T2-weighted fast spin echo sequence, and postcontrast and postcontrast fat-suppressed T1-weighted gradient spin echo sequence were obtained by ACS and PI. Acquisition time, signal-to-noise ratio, contrast-to-noise ratio, and image quality of both sets of MRI simulation images were compared. Image quality analysis was scored with lesion detection, margin sharpness of lesions, artifacts, and overall image quality using the 5-point Likert scale. Moreover, tumor target volume acquired from fusion images of simulation computed tomography with simulation MRI by ACS and from fusion images by PI were compared. Dice similarity coefficient of gross tumor target between fusion images by ACS and those by PI were also measured.

Results

Acquisition time of MRI simulation by ACS was significantly shorter than that by PI, whether for the time of individual sequence or the total acquisition time (P < .05 for all). The mean total acquisition time by PI (694.78 ± 16.85 seconds) was significantly less after using ACS (378.50 ± 10.05 seconds), with a mean reduction ratio 45.52%. Signal-to-noise ratio, contrast-to-noise ratio values and qualitative image scores (lesion detection, margin sharpness, artifacts, and overall image quality) were almost comparable between ACS and PI. Mean tumor target volume of both primary tumors and metastatic lymph nodes acquired from fusion images by ACS were also comparable to those from fusion images by PI (P > .05 for all). Mean Dice similarity coefficient values for primary tumors and metastatic lymph nodes were both close to 1.

Conclusions

Compared to PI, ACS can significantly accelerate MRI simulation for head and neck cancer radiation therapy without compromising image quality and degrading the guidance role of tumor target delineation.