Medical image analysis methods for anatomical surface reconstruction using tracked 3D ultrasound

Abstract

The thesis focuses on a study of techniques for acquisition and reconstruction of surface data from anatomical objects by means of tracked 3D ultrasound. In the context of the work two experimental scanning systems are developed and tested on both artificial objects and biological tissues. The first system is based on the freehand ultrasound principle and utilizes a conventional 2D ultrasound transducer coupled with an electromechanical 3D position tracker. The main properties and the basic features of this system are discussed. A number of experiments show that its accuracy in the close to ideal conditions reaches 1.2 mm RMS. The second proposed system implements the sequential triggered scanning approach. The system consists of an ultrasound machine, a workstation and a scanning body (a moving tank filled with liquid and a transducer fixation block) that performs transducer positioning and tracking functions. The system is tested on artificial and real bones. The performed experiments illustrate that it provides significantly better accuracy than the freehand ultrasound (about 0.2 mm RMS) and allows acquiring regular data with a good precision. This makes such a system a promising tool for orthopaedic and trauma surgeons during contactless X-ray-free examinations of injured extremities. The second major subject of the thesis concerns development of medical image analysis methods for 3D surface reconstruction and 2D object detection. We introduce a method based on mesh-growing surface reconstruction that is designed for noisy and sparse data received from 3D tracked ultrasound scanners. A series of experiments on synthetic and ultrasound data show an appropriate reconstruction accuracy. The reconstruction error is measured as the averaged distance between the faces of the mesh and the points from the cloud. Dependently on the initial settings of the method the error varies in range 0.04 - 0.2% for artificial data and 0.3 - 0.7 mm for ultrasound bone data. The reconstructed surfaces correctly interpolate the original point clouds and demonstrate proper smoothness. The next significant problem considered in the work is 2D object detection. Although medical object detection is not integrated into the developed scanning systems, it can be used as a possible further extension of the systems for automatic detection of specific anatomical structures. We analyse the existent object detection methods and introduce a modification of the one based on the popular Generalized Hough Transform (GHT). Unlike the original GHT, the developed method is invariant to rotation and uniform scaling, and uses an intuitive two-point parametrization. We propose several implementations of the feature-to-vote conversion function with the corresponding vote analysis principles. Special attention is devoted to a study of the hierarchical vote analysis and its probabilistic properties. We introduce a parameter space subdivision strategy that reduces the probability of vote peak omission, and show that it can be efficiently implemented in practice using the Gumbel probability distribution.