Segmenting Dermoscopy Images Using Variational Level Sets With An External Force Based On Supervised Learning

Abstract

Melanoma is one of the most common and deadliest form of skin cancer. The world- wide increase in mortality rate because of the incidence of malignant melanoma leads to the need of the early detection of melanoma. A lot of research is done in this eld by using computer vision and pattern recognition techniques for the diagnosis of melanoma cancer. Keeping in view the earlier research, this thesis proposed a Computer Aided Diagnosis (CAD) System for detection of melanoma by proposing novel image processing approaches. The main focus of the thesis is to develop a new numerical algorithm for segmenta- tion of region of interest in melanoma image which is a very important step in any image processing approach. A new algorithm is proposed by integration of Gaus- sian mixture model with the variational level sets using supervised learning. By calculating dice similarity coe cient (DSC) and F-measure, the performance of an algorithm is compared with the segmented masks given by the physician. After- wards, the texture features are used for classi cation to validate the segmentation results. Experimental results using Support Vector Machines show the performance of algorithm by comparing it with ground truth (manual annotations).