Accelerated ultrashort echo time quantitative magnetization transfer (UTE-qMT) imaging of macromolecular fraction (MMF) in cortical bone based on a self-attention convolutional neural network

• Published in: Magnetic resonance imaging Vol. 121; p. 110405
• Main Authors: Du, Kevin, Tang, Harry, Athertya, Jiyo, Wang, Yidan, Hu, Megan, Wang, Avery, Jerban, Saeed, Shin, Soo Hyun, Ma, Yajun, Chung, Christine B., Chang, Eric Y.
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier Inc 01.09.2025
• Subjects: Adult; Algorithms; Bone; Bone Diseases, Metabolic - diagnostic imaging; Convolutional Neural Networks; Cortical Bone - diagnostic imaging; Female; Humans; Image Processing, Computer-Assisted - methods; Magnetic Resonance Imaging - methods; Middle Aged; MMF; MRI; Neural Networks, Computer; Osteoporosis - diagnostic imaging; SAT-net; Tibia - diagnostic imaging; UTE-qMT; Young Adult; SAT-net; OP; UTE-qMT; Bone; MRI; MMF
• DOI: 10.1016/j.mri.2025.110405
• ISSN: 0730-725X; 1873-5894

To combine ultrashort echo time quantitative magnetization transfer (UTE-qMT) imaging with a self-attention convolutional neural network (SAT-Net) for accelerated mapping of macromolecular fraction (MMF) in cortical bone. This institutional review board-approved study involved 31 young female subjects (young control, <45 years) and 50 postmenopausal subjects (6 normal (old control), 14 with osteopenia (osteopenia group), and 30 with osteoporosis (OP group)). After written informed consent was obtained from each subject, 15 UTE-qMT images of the tibial midshaft were acquired with three saturation powers (500°, 1000°, and 1500°) and five frequency offsets (2, 5, 10, 20, and 50 kHz) for each power to estimate the baseline MMF using a two-pool model. The densely connected SAT-Net model was used to predict bone MMF maps based on seven evenly distributed UTE-qMT images, which were well separated in terms of MT powers and frequency offsets (namely 5 and 20 kHz for 500° and 1500°, and 2, 10, 50 kHz for 1000°). Errors relative to the baseline MMF were calculated. Linear regression was used to assess the performance of the SAT-Net model. The mean MMF values for different groups were calculated. Conventional two-pool modeling of seven evenly distributed UTE-qMT input images shows a significant relative error of ∼34 %. In comparison, the SAT-Net model accurately predicted MMF values for the tibial midshafts of 81 human subjects with a high correlation (R2 = 0.97, P < 0.0001) between the baseline and predicted values. The SAT-Net model accelerated UTE-qMT data acquisition by 2.1-fold, with relative errors in MMF mapping less than 2.4 %. The average MMF values were 46.10 ± 13.25 % for the young control group, 40.03 ± 2.56 % for the old control group, 31.22 ± 13.18 % for the osteopenia group, and 22.53 ± 8.12 % for the OP group. While it is difficult to accelerate MMF mapping in bone using conventional two-pool modeling, the SAT-Net model allows accurate MMF mapping with a substantial reduction in the number of UTE-qMT input images. UTE-qMT with SAT-Net makes clinical evaluation of bone matrix possible.