Automated Detection and Segmentation of Bone Metastases on Spine MRI Using U-Net: A Multicenter Study
Authors
Dong Hyun Kim, Jiwoon Seo, Ji Hyun Lee, Eun-Tae Jeon, DongYoung Jeong Hee Dong Chae, Eugene Lee, Ji Hee Kang, Yoon-Hee Choi, Hyo Jin Kim and Jee Won Chai
This study developed and evaluated a deep learning model using U-Net for automated segmentation and detection of bone metastasis on spinal MRI. Data included 662 MRI series from 302 patients for training and internal testing, and 49 MRI series from 20 patients for external testing. The MRI sequences used were non-contrast T1-weighted, contrast-enhanced T1-weighted Dixon fat-only, and contrast-enhanced fat-suppressed T1-weighted images. Seven models using 2D and 3D U-Nets were tested. The best performance was achieved by the 2D U-Net T1 + CE model with a Dice coefficient of 0.699 and pixel-wise recall of 0.653. For per-lesion sensitivity, the T1 + CE model outperformed others with 0.828 for internal and 0.857 for external tests. Radiologists showed a mean per-lesion sensitivity of 0.746. Segmentation and detection were refined using AVIEW Research software. The proposed deep learning models demonstrated high diagnostic performance for detecting bone metastases on spinal MRI.