Feasibility of a lumbar CT-to-MRI image translation framework using a conditional CycleGANs for fracture stage determination in osteoporotic compression fractures.
Authors
Affiliations (7)
Affiliations (7)
- Department of Neurosurgery, Inje University Ilsan Paik Hospital, Goyang-si, Republic of Korea.
- Department Bio-mechanical Engineering, Korea University, Seoul, Republic of Korea.
- Department of Radiation Oncology, Asan Medical Center, Seoul, Republic of Korea.
- Department of Neurosurgery, Spine Center, Nanoori Gangnam Hospital, Seoul, Republic of Korea.
- Department of Orthopedic surgery, Suncheon Joongang hospital, South Korea, Suncheon, Republic of Korea.
- Department of Radiology, Inje University Ilsan Paik Hospital, Goyang-si, Republic of Korea.
- Department of Neurosurgery, Inje University Ilsan Paik Hospital, Goyang-si, Republic of Korea. [email protected].
Abstract
Magnetic resonance imaging (MRI) provides a reliable indicator of fracture stage. However, MRI scans can be expensive, time-consuming, and not readily accessible to all patients. With recent advances in artificial intelligence, translating lumbar CT images into MRI images has become feasible using cycle generative adversarial network (CycleGANs) based approaches. The aim of this study was to evaluate the feasibility of a model that translates lumbar CT images into MRI images (T1-weighted contrast-enhanced sequences) to predict the stage of osteoporotic compression fracture (OCFs) using conditional CycleGANs. A total of 652 patients diagnosed with OCFs who underwent both CT and MRI within 2 weeks were retrospectively collected. Of these, 552 patients were used for training, and an independent set of 100 patients was reserved for testing. CT and MRI volumes were aligned using segmentation-based rigid registration. The proposed framework consisted of two stages: generation of vertebra-specific acute fracture maps from CT images using a 3D ResNet-based model, followed by CT-to-MRI image translation using a 3D conditional CycleGANs with acute stage conditioning. Performance was evaluated using vertebra-level classification metrics, image similarity metrics, and qualitative assessment by five experienced readers, including three spine neurosurgeons, one orthopedic spine specialist, and one radiologist. Fractures were most common at the thoracolumbar junction. A total of 800 vertebral bodies were analyzed. Of 678 non-acute OCFs, 667 were correctly classified and 11 were misclassified as acute; of 122 acute OCFs, 106 were correctly identified and 16 were misclassified as non-acute. Acute fracture classification achieved a precision of 0.906, recall of 0.869, and F1-score of 0.887. Synthetic MRI images showed modest quantitative agreement with real MRI (Normalized MAE 0.094 ± 0.024; PSNR 18.7 ± 1.7 dB; SSIM 0.38 ± 0.04). Qualitative evaluation demonstrated high similarity in vertebral body morphology at the fracture level, with mean similarity scores of 87.2% for the conditional CycleGAN and 85.3% for the conventional CycleGAN. This study demonstrates the feasibility of a lumbar CT-to-MRI image translation framework using a conditional CycleGAN for determining fracture stage.