loading page

CNN Detection of Vertebral Bone Segmentation and Non-Traumatic Vertebral Compression Fractures from Computed Tomography (CT) Images
  • Murat Türkmen
Murat Türkmen
Istanbul University-Cerrahpasa

Corresponding Author:[email protected]

Author Profile


Vertebral compression fractures are a more common pathological condition with prolonged life expectancy. The decrease in bone mass with age is effective in the formation of non-traumatic vertebral compression fractures. Since non-traumatic vertebral compression fractures are usually asymptomatic for a long time, early detection in this period offers a reduction in comorbidities and the possibility of more effective treatment. Deep learning methods offer effective solutions for the detection, classification, diagnosis, and segmentation of pathological conditions with high accuracy and sensitivity in the field of health. In recent years, deep learning methods have been used more frequently in the detection of non-traumatic vertebral compression fractures and vertebral corpus segmentation research. In this study, a unique data set is presented for the application of deep learning methods by processing the raw Computed Tomography (CT) image data of the patients. The original data set was created from retrospective CT images of patients who received radiological diagnosis with CT in Istanbul University-Cerrahpasa, Cerrahpasa Faculty of Medicine, Department of Radiology. The original dataset consisted of 197 people, 100 patients with non-traumatic vertebral compression fractures and 97 patients without. Vertebral compression fracture radiological diagnoses were added to the data set with a CT report. A total of 197 people have 118200 cross-section images in DICOM format. Images are enhanced by applying the wiener filter. Segmentation of 6301 vertebrae was performed using the U-Net network at 100% dice overlap index score. 593 features of vertebral fractures confirmed by the report were extracted using the radiomics method. 537 features were selected with the logarithmic lambda method. Patients were classified with 86.7% accuracy using the Convolutional Neural Network (CNN) classification model. The classification result was evaluated by creating roc-auc, loss, and accuracy graphs.