Due to its ease of use, portability, low cost, real-time response and absence of ionising radiation, ultrasound (US) imaging could potentially be an important tool for non-invasive diagnostic imaging in orthopaedics. Unfortunately, nonlinear characteristics of ultrasound, low signal-to-noise ratio and speckle make it difficult to accurately and reliably determine the location and shape of the bone surface. Recently, local phase-based image processing methods, named phase symmetry (PS), have been shown to perform very well at locating bone surfaces in ultrasound images, with reported accuracies of better than 0.4mm. The local phase features are extracted by filtering the B-mode US image in the frequency domain with a Log-Gabor filter. Although successful results were achieved, accurate localization is highly affected by the choice of filter parameters. Recently, our group proposed a method of automatically selecting the scale, bandwidth and orientation parameters of Log-Gabor filters. Previously, we showed our first clinical results using local phase information to identify distal radius fractures from B-mode US images using automatically selected filter parameters. The objective of the current study was to determine if the proposed automatic parameter selection method could produce accurate pelvic bone surface shapes in a live clinical setting. CT scans were obtained as part of normal clinical care from ten patients admitted to Vancouver General Hospital for pelvic fractures. A ‘gold standard’ bone surface was computed from the CT scan. After obtaining informed consent, we performed an additional US scan using a commercially-available real-time scanner (Voluson 730, GE Healthcare, Waukesha, WI) with a 3D US transducer. The PS bone surfaces were extracted from the US scans using the empirical Log-Gabor filter parameters and optimised Log-Gabor filter parameters. The bone surfaces on CT were extracted using a standard thresholding approach that minimises the intra-class variance. The US images were then registered to the CT images using a feature-based rigid registration algorithm with manual landmarking. The quality of the resulting surface matching was evaluated by computing the root mean square distance between the two surface representations. The average fiducial registration error was 0.31mm (SD 0.25mm). The average surface fitting error (SFE) was 0.72mm (SD 1.24 mm) for PS surfaces extracted using empirical filter parameters and 0.41mm (SD 0.44 mm) using the optimized filter parameters. In this study, we have demonstrated that our automatic filter parameter selection process can be applied successfully to a bone surface extraction task on 3D US images acquired under clinically realistic conditions. The accuracy of the resulting bone surface is excellent, with an average discrepancy relative to a CT standard of well under a millimeter. This level of accuracy is likely to be sufficiently good for a number of important surgical tasks, including CT to US registration.
