8

Conclusions


 

This chapter contains four principal conclusions that have been drawn from the work carried out. The original aim of this work is to improve the precision of wound measurement by removing the subjective elements of human observation and manual dexterity. The properties of most wound images prevent many deterministic image processing methods from offering a viable alternative to the established manual methods, and instead attention has been focussed on active contour models to provide a mechanism for refining manual wound area measurements.

 

8.1 Manual Delineation Performance

The manual wound measurement results of Chapter 4 show that not all of the delineators involved in the trial measure the same wound area. This is due to differences of opinion concerning which parts of an image belong to a wound. Also, significant differences in measurement precision show that some delineators generally achieve a higher level of repeatability in their measurements. The precision level varies with the delineator, the wound image itself, and the size of the display of the image on a monitor. Images of wounds that appear well defined have relatively strong edge strength at the wound boundary. The last variable, the size of image displayed, means that the precision of measurements made on smaller wounds improves with magnification.

 

8.2 Algorithm Performance

Active contour models are appropriate for the wound measurement problem because they require an initial estimate of the final boundary, which can be provided by a trained delineator. Thus the solution is obtained by local modification of the contour, without regard to general properties of the enclosed region or of the surrounding tissue. Four different formulations of active contour model algorithms have been tested. The TN algorithm has proved to be the most appropriate single algorithm for improving wound measurements:

The GA and MG algorithms, whilst generally inferior in performance to the TN algorithm supersede it when the image gradient information becomes too ambiguous or when the initial filter scale is too high. In such cases the gray level term of the MG algorithm replaces the lost information, and the GA algorithm benefits from retaining the initial contour in the iteration equations. The setting of contour regularisation parameters to apply across a wide range of images causes precision loss in the GA and TN algorithms. This requires a method for adapting the parameters to each image without compromising the stability of the solution. The parameter-less MX algorithm, which is adaptive in this sense, does not perform as well as the TN algorithm.

Automation of the algorithms has not been achieved because it is considered that not enough reliable information generally exists in wound images in general to allow automatically initialised contours or region growing contours to be deformed to a reliable solution.

 

8.3 General Statement

This thesis claims the following innovations as part of its original contribution to knowledge:

The algorithms described in this work, their implementation and testing has resulted in a useful wound measurement tool. The work on parameter setting in Chapter 5 enables the algorithms to operate with little end-user intervention. Thus they may be used in a clinical setting by a physician and do not require technical or expert intervention. The algorithms produce improved precision area measurements, the only human effort required being an initial, approximate delineation of the wound boundary using a mouse.

The image and delineator-conditional ability of the algorithms to improve precision should be weighed carefully on a case-by-case basis, so that when complex or poorly defined wounds are measured or where poorly posed and poorly illuminated video capture is apparent the expectation for improved precision may not be fulfilled.

Since so many cases (40%) produce bias errors, a plan for wound measurement that accounts for this factor is therefore recommended:

 

8.4 Limitations of this Work

There are many possible sources of error in video measurements (Plassmann 1992). In terms of experimental analysis the most obvious missing component of the work is an investigation into the errors arising from taking measurements from different images of the same wound taken at the same visit. The measurement of image area has concentrated on the 2-D projection of the object onto the camera plane. Further work is necessary therefore to estimate the sum contribution of these error sources and weigh their effect upon the reliability of measurements.

There is much scope for further investigation in the area of contour initialisation. Chapter 2 discussed the various methods reported in practice and introduced some ideas. However, the algorithms reported in this thesis have not employed such methods due to the requirement that they work across a broad range of wound images. This could well involve a large amount of work determining amongst other things, the region, both inside and outside of the wound boundary, within which a contour can be placed and successfully converge to a solution with an error on a par with the results quoted in Chapter 6.

 

8.5 Recommendations for Further Work

Since bias errors between manually defined areas and algorithm defined areas are the main source of error it is recommended that a further study be performed which tracks the progress of a set of wounds. The same procedures for measurement as described in this work may then be used to determine whether or not the bias errors can be wholly or partly ignored, or whether they introduce a further degree of variability. If the latter case is found to be true, then a systematic error in one set of measurements will become essentially a random error when considered over many such sets of measurements, i.e. the bias between manual and algorithmic measurements will be unpredictable from one wound inspection to the next.

A further idea would be to employ the multiple contour principle of Gunn and Nixon (1994) so that several starting estimates would be supplied for a single measurement. Instead of these contours being mutually independent, the multiple contour formulation would enforce a degree of similarity between results, thereby reducing variability and thus increasing precision. Furthermore, it may be possible, where more than two contours are used, to consider energies for contour segments that are in closer agreement and to exclude any significantly different contour segments. The multiple-contour configuration should not require so much dependence upon scale descent techniques to improve precision and thus should not suffer so much from the effect of bias.


Title Page

4 Manual Delineation

Appendix A

Contents

5 Parameter Setting

Appendix B

1 Introduction

6 Results

Appendix C

2 Literature Survey

7 Discussion

Colour Plates

3 Algorithm Development

8 Conclusion

References