Developing improved algorithms for detection and analysis of skin cancer
- Publication Type:
- Issue Date:
Malignant melanoma is one of the deadliest forms of skin cancer and number of cases showed rapid increase in Europe, America, and Australia over the last few decades. Australia has one of the highest rates of skin cancer in the world, at nearly four times the rates in Canada, the US and the UK. Cancer treatment costs constitute more 7.2% of health system costs. However, a recovery rate of around 95% can be achieved if melanoma is detected at an early stage. Early diagnosis is obviously dependent upon accurate assessment by a medical practitioner. The variations of diagnosis are sufficiency large and there is a lack of detail of the test methods. This thesis investigates the methods for automated analysis of skin images to develop improved algorithms and to extend the functionality of the existing methods used in various stages of the automated diagnostic system. This in the long run can provide an alternative basis for researchers to experiment new and existing methodologies for skin cancer detection and diagnosis to help the medical practitioners. The objective is to have a detailed investigation for the requirements of automated skin cancer diagnostic systems, improve and develop relevant segmentation, feature selection and classification methods to deal with complex structures present in both dermoscopic/digital images and histopathological images. During the course of this thesis, several algorithms were developed. These algorithms were used in skin cancer diagnosis studies and some of them can also be applied in wider machine learning areas. The most important contributions of this thesis can be summarized as below: - Developing new segmentation algorithms designed specifically for skin cancer images including digital images of lesions and histopathalogical images with attention to their respective properties. The proposed algorithm uses a two-stage approach. Initially coarse segmentation of lesion area is done based on histogram analysis based orientation sensitive fuzzy C Mean clustering algorithm. The result of stage 1 is used for the initialization of a level set based algorithm developed for detecting finer differentiating details. The proposed algorithms achieved true detection rate of around 93% for external skin lesion images and around 88% for histopathological images. - Developing adaptive differential evolution based feature selection and parameter optimization algorithm. The proposed method is aimed to come up with an efficient approach to provide good accuracy for the skin cancer detection, while taking care of number of features and parameter tuning of feature selection and classification algorithm, as they all play important role in the overall analysis phase. The proposed method was also tested on 10 standard datasets for different kind of cancers and results shows improved performance for all the datasets compared to various state-of the art methods. - Proposing a parallelized knowledge based learning model which can make better use of the differentiating features along with increasing the generalization capability of the classification phase using advised support vector machine. Two classification algorithms were also developed for skin cancer data analysis, which can make use of both labelled and unlabelled data for training. First one is based on semi advised support vector machine. While the second one based on Deep Learning approach. The method of integrating the results of these two methods is also proposed. The experimental analysis showed very promising results for the appropriate diagnosis of melanoma. The classification accuracy achieved with the help of proposed algorithms was around 95% for external skin lesion classification and around 92 % for histopathalogical image analysis. Skin cancer dataset used in this thesis is obtained mainly from Sydney Melanoma Diagnostic Centre, Royal Prince Alfred Hospital. While for comparative analysis and benchmarking of the few algorithms some standard online cancer datasets were also used. Obtained result shows a good performance in segmentation and classification and can form the basis of more advanced computer aided diagnostic systems. While in future, the developed algorithms can also be extended for other kind of image analysis applications.
Please use this identifier to cite or link to this item: