Mixture Models and Markov Random Fields |
| General |
This is a 4th year Software Engineering Honours project. This research project was supervised by Dr. David Squire |
| Project Information |
|
Australia is a country known for its dry and sunny conditions and a high rate of skin cancer. There are many different types of skin cancer, and of the most dangerous of them all is Melanoma. Melanoma is caused from unprotected high exposure to UV radiation. Commonly, people with fair skin or skin that contains lots of freckles and moles are at high risk. Melanoma has claimed the lives of many people. It is estimated annually that over 1000 people in Australia alone die from melanoma, 7700 and 1600 people in the US and the UK respectively. The annual rate of melanoma cases worldwide is steadily increasing at around six percent. It is an extremely alarming rate. Fortunately, early detection of this skin cancer can improve the cure rate, and in most cases a 100% cure rate is possible. However, if the melanoma is not detected early and the lesion is more than three millimetres deep, the chance of survival is 59%. Most general practitioners have less experience in the full range of melanoma forms, and because of this, many melanoma cases are not diagnosed properly. In recent years, there has been much research and development into the area of skin lesion analysis. The majority of skin lesion analysis involves taking image samples of the suspicious areas from a human body and then further analysing the images for signs of skin cancer. Surface Microscopy and Dermoscopy images are two of the main images commonly used by dermatologists. One of the most important aspects of studying dermoscopy images is the detection of skin cancer, particularly for signs of the Malignant Melanoma lesion. The fundamental task in lesion detection is to separate healthy skin from the skin lesion. The traditional pre-processing for the detection of melanoma involves manually drawing borders around the lesion areas on the image. Although it is a simple method, it is time consuming and very subjective, as each individual dermatologist has a different level of experience. Current research methods tackle this problem through the use of image segmentation and edge detection, however the processes involved to produce accurate results are still far from perfect. Techniques such as analysing the image contours, colour, and texture for segmentation, while symmetry and the regularity of the lesions for subsequent analysis have been tried, and yet the methods and processes involved are still lacking overall accuracy. Other new techniques must be considered for the benefit of the current problem. This is the goal of my research project, to study effective segmentation models based on Markov Random Fields and Gaussian Mixture Models. |