This thesis examines the analysis of digitised optical skin cancer images. Its aims are to develop a new method of skin lesion boundary detection that meets the need for reliable automated boundary detection; to develop and assess shape analysis methods that are suitable for skin lesion diagnosis and can be related to human shape perception to allow application of expert knowledge; and to develop methods of creating simulated skin lesion images to allow study of the behaviour of boundary detection methods.
Methods for analysing lesion shape are presented and tested on real and synthesised shapes. The effect of boundary noise on shape analysis is investigated and shown to increase the bulkiness and textural and structural fractal dimensions. It is shown that the necessity of using high resolution shapes and the effect of noise on fractal dimension measurement indicates that it may not be suitable for assessing lesion shape.
A new method for locating and isolating a lesion within a skin image is presented, which provides an image containing the lesion and surrounding skin, but excluding background objects, with an indication of the lesion's size. This method correctly identified all of the lesions in a test set. Its extension to images containing multiple lesions is also discussed.
An edge focusing algorithm for skin lesion boundary detection, which uses either
Laplacian of Gaussian (LoG) or Canny edge detection, is presented and tested on real images. This is a new application of edge focusing, which uses new methods to control the boundary during focusing and to select the output boundary, using image contrast.
By combining this algorithm with the process of isolating lesions, a system that can find lesion boundaries in a range of images is produced. Testing using real and verification images indicate that it is capable of working on a wide range of images. A new method for synthesising simulated skin lesion images is presented and used to assess the edge focusing algorithm using an area based comparison method. Using parameters, derived from real images, the lesion location and isolation method correctly identified all the simulated lesions. The performance of the LoG and Canny edge focusing algorithms is shown to decrease as the lesion boundary becomes less distinct.
The development of a computer based tool to perform or assist in the diagnosis of skin cancer and how the research presented in this thesis would be incorporated into such a tool is discussed. Further research into the measurement techniques, required to obtain diagnostic and prognostic information, is outlined and the use of this information in
providing diagnosis and prognosis is examined.