Introduction: Rapid and accurate diagnosis of an osteoporotic vertebral compression fractures (OVCF) is crucial for appropriate treatment. However, in some cases , timely diagnosis and treatment are hindered by shortages of medical personnel and testing equipment. Deep learning-based artificial intelligence (AI) can solve such situations. This study aimed to assess the efficacy of deep learning-based AI in detecting OVCF and calculating compression rates on X-ray images .
Methods: Between January 2019 and December 2023, 100 patients with suspected vertebral compression fractures, including 30 without fracture, were enrolled. Patients underwent anteroposterior and lateral spine X-rays , followed by CT or MRI within 5 weeks. BMD was assessed within a year of the X-ray. Each patient's plain radiograph was assigned a unique identification number, and diagnoses and compression rate calculations were based solely on these X-rays. To evaluate the AI model in diagnosing osteoporotic vertebral fractures and calculating compression ratios, results were compared with those of two orthopedic surgeons (with 1 and 14 years of experience), one orthopedic resident, and one general physician.
Results: AI demonstrated a diagnostic accuracy rate of 0.72, while the orthopedic surgeon with 14 years of experience achieved 0.85, surgeon with 1 year of experience 0.84, orthopedic residents 0.82, and general physicians 0.80. Statistically significant differences in diagnostic accuracy were observed between AI and both orthopedic surgeons (p = 0.01) and the orthopedic resident (p = 0.04), but not with the general physician ( p = 0.15). Accuracy comparisons among the physicians themselves were not statistically significant. Pearson correlation coefficients for compression rates showed strong positive correlations between AI and each group of physicians, as well as among the physicians themselves (p < 0.01).
Discussion: The result suggests that the current morphometric analysis AI model is not sufficient to be used independently for diagnosing compression fractures. However, it shows a certain level of accuracy that can serve as a supportive tool to help physicians identify vertebral compression fractures. Additionally, the AI model can be helpful in situations where the number of doctors is limited and the workload is excessive. The ability to immediately and independently diagnose OVCFs using plain radiographs will significantly impact clinical image diagnoses. This will improve the prognosis and survival rates of OVCF patients. To achieve this capability, the deep learning-based AI model needs further development.
Most previous studies have used morphology analysis for recognizing OVCF. Our study used a deep learning-based AI morphometric model compared to prior studies. The difference between the morphology and morphometric models lies in how they detect vertebral compression fractures. The morphology analysis AI model first estimates and compares the normal vertebra shape with the patient’s vertebra shape. However, the morphometric analysis AI model first calculates the compression fracture rate of each vertebra in the plain radiograph. Vertebrae with a high compression fracture rate (more than 20%) are then diagnosed as vertebral compression fractures. This different method of diagnosing OVCF is the reason for the lower accuracy of the morphometric analysis AI model compared to the morphology analysis model.