Melanoma Detection Using Image Processing and Computer Vision Algorithms
Melanoma is the most serious form of skin cancer and is estimated to be the 19th most frequently occurring type of cancer worldwide, with approximately 232,000 new cases diagnosed in 2012. It is widely accepted that early diagnosis of melanoma significantly reduces morbidity, mortality and the cost of medication. Computer-aided systems can be applied for a quantitative and objective evaluation of pigmented skin lesions to assist the clinical assessment process. Increasing innovation in non-invasive methods can be of significant help in the early detection of malignant melanoma, thus minimising the need for biopsies. The initial step is to analyse and develop efficient algorithms for melanoma detection. This thesis is mainly focused on two main areas: a) developing an efficient lesion border detection algorithm, and b) developing an efficient classification system. For lesion border detection, several edge detection techniques are evaluated. We implemented a basic border detection algorithm on the ZYNQ-7000 System-on-Chips, which suggests a proper portable vision system could be designed for early detection of melanoma with high resolution and performance. A semi-automatic algorithm consisting of eight steps is proposed for detecting the borders of skin lesions in clinical images. Using this approach, the user selects a small patch of the lesion to specify the foreground lesion area. The results show that the proposed method achieved the accuracy of 89.32%. We present a multi-layer feed-forward deep neural networks (DNN) as a preferred lesion segmentation and recognition method. The algorithm can detect lesion borders without using any pre-processing algorithms; however, a pre-processing step hair removal and illumination correction has been essential in the previous systems. In order to develop a classification system, we investigated two different approaches: a) using hand-engineered feature data that are extracted from the segmented lesion and b) using a deep learning method which learns features automatically from the original images. The feature extraction algorithms that are used in this study are shape, colour and texture features. Correlation-based feature selection method is applied for feature selection. A performance evaluation of several supervised classifiers are discussed based on different feature sets. Two novel cascade classification architectures are proposed to improve accuracy. The second proposed cascade classifier achieved an overall accuracy of 83.3%, sensitivity of 85.1%, specificity of 80% and ROC area of 90% using ten-fold cross-validation. Finally, we present a multi-layer DNN to distinguish melanoma from benign nevi as our preferred method for classification. Our preliminary work shows that networks trained with no pre-processed and segmented images, using directly learned features instead of applying feature extraction; achieved an average accuracy of 72.53%. However, a larger dataset and more investigations are required to train a better classifier.