Computer Aided Diagnosis System for Melanoma Lesion Detection

Detection of skin cancer in the earlier stage is very critical. Nowadays, skin cancer is seen as one of the most hazardous form of cancers found in humans. The most common types of skin cancers are Basal Cell Carcinoma (BCC), Squamous Cell Carcinoma (SCC) and Melanoma. Among these melanoma is the deadliest type of skin cancer. The detection of Melanoma cancer in early stage can be helpful to cure it. Computer vision plays an important role in medical image diagnosis and this has been proved by many existing systems. A computer aided skin image diagnosis system has a significant potential for screening and prognosis. It is utmost important in countries where unaided visual diagnosis system is the practice and where there is insufficient number of dermatologists. This thesis presents a computer aided diagnosis system for the detection of melanoma skin cancer using a novel mathematical scheme. The proposed method uses a holistic representation of skin color images to extract useful features for use in effective segmentation of melanoma lesions. The segmentation scheme is preceded by a processing stage composed of noise filtering color space transformation. Vertex component analysis and principal component analysis are integrated to form a hybrid approach for unmixing and feature dimension reduction. An optimized feature selection technique is also used to obtain the best achievable performance in effectively detecting the lesions. Support vector machine (SVM) along with geometrical and color feature threshold values are integrated to make effective detection of melanoma lesions. The effectiveness of the developed scheme to classify lesions into benign, suspicious and malignant melanoma is found to be promising. The proposed scheme has been tested on images taken from standard dermoscopy image databases and achieved 96.4% sensitivity, 99.4% specificity, and 97.9% overall accuracy for pixel based classification of melanomas while it achieved 98% sensitivity, 100% specificity, and 98.6% accuracy for image based classification justifying it"s great promises.