Automatic segmentation of vertebral interfaces in echographic images

Quantitative ultrasound (QUS) methods for osteoporosis diagnosis potentially provide information about the bone quality and its elastic properties. In this context, a novel ultrasound-based method for spinal and femoral densitometry was developed by our research group. In order to maximize its accuracy, it is very important to properly detect the bone interfaces that will be analyzed as regions of interest (ROIs). A fully automatic segmentation algorithm was developed to select lumbar vertebral interfaces in echographic images and its actual accuracy was assessed in the present work by means of a visual checking carried out by an expert operator. Abdominal US scans of lumbar spine (from L1 to L4) were performed on 100 female subjects (60.5±3.0 years old) with different ranges of body mass index (BMI) (25.8±4.6 kg/m2). During each US scan, 100 frames of radiofrequency (RF) data were stored on a PC hard disk for offline analysis. The operator scanned each vertebra, moving the probe to the next vertebra after 20 seconds. For each acquired RF data frame, the implemented algorithm generated a sectorial echographic image and, if a vertebral interface was detected, it was highlighted on the saved image. The validation procedure was performed by an expert operator previously trained to detect the "optimal" vertebral interfaces for osteoporosis diagnosis. Results showed that the segmentation algorithm had a high specificity (93.4%), which reached its maximum on subjects with BMI < 25 kg/m2 (94.2%), thus avoiding the selection of false vertebral interfaces and allowing a good accuracy of osteoporosis diagnosis.