Medical Application: Quantitative Analysis of Biomedical Images

Medical Application:
Quantitative Analysis of Biomedical Images

Introduction
The use of in vivo imaging techniques is becoming increasingly widespread as reduced costs and technological advancement broadens its application and appeal. There is a requirement in the pharmaceuticals industry and medical research in general to increase the efficiency of drug trials and disease progression studies by using fewer subjects in shorter trial periods. In constrast, medical diagnosis and surgical planning, which usually involve the use of a single image of an individual, require more effective diagnosis and patient management. While in vivo imaging may begin to address these requirements, the image analysis stage presents a bottleneck in in the acquisition of useful information.

This bottleneck arises from a large volume of data for which the accuracy of analysis needed may be high. Current image analysis techniques are too cumbersome and slow for this task. This is because these techniques require a great deal of user intervention. The goal of this project is to provide a generic solution to the task of quantitative image analysis. This solution should be an integrated system, free from the complexities of a system design for the machine vision or image processing expert.

Approach
The quantitative analysis of medical images relies upon the
segmentation of relevent regions of interest from the image data. Data-driven segmentation techniques usually require some user interaction, especially if the data is noisy or structural boundaries are incomplete.We address this problem using Active Shape Models (ASMs), flexible templates incorporating shape constraints:

A model of the object is built by an expert from a number of training images using identified landmarks. This produces a `flexible template' with a small number of controlling parameters which may then generate new examples of the object.

A search is performed employing the model in order to locate the anatomical region of a new image which most likely to correspond to the modelled object.

A plausible result is guaranteed by incorporating such a priori knowledge of the objects of interest. This approach is well suited to the interpretation of medical images.

Building a Model
Within the 3D image volume, the shape of an object is initially defined as a set of boundaries on 2D slices, consisting of an ordered hierarchy of topology objects - vertices, edges and chains of edges. Two or more edges may meet at a common vertex, as for example in the definition of the growth plate of the femur in the above MR image of a human knee. The topological definitions on a series of individual slices are joined to make a surface by triangulation.

The individual topological entities are assigned to parts of the anatomical structure being defined. The parts are given descriptive names and exist within a graph structure which allows the definition of part/subpart hierarchies. Several hierarchies of anatomical parts may be defined within a single image.

System Requirements
To cope with short analysis runs, the new system must reduce the time overhead of model-building by `learning-on-the-job'. Initially, the standard methods of object defintion may be used to produce anatomical structures of interest. As the analysis progresses, the system will become more intelligent, proposing automatic segmentation based upon the use of incrementally constructed flexible template models. At this stage, segmentation errors may be corrected interactively. As the examples are accumulated, the need to make corrections will steadily decrease as the models become better trained. Ultimately, this leads to completely automatic segmentation and measurement.

Acknowledgements
Funding for this project was provided by the EPSRC IMV initiative and involves collaboration between Zeneca Pharamaceuticals, IBM and the WIAU.

Alan Brett:

Research Courses Clinical Radiology Industrial liaison Contact Us Join Us Search Home

Top Page