Investigations on performance analysis of classifiers with deep learning features for classification of melanoma from dermoscopy images

Abstract

Melanoma is accounted as a rare skin cancer responsible for a huge newlinemortality rate. However, various imaging tests can be used to detect the newlinemetastatic spread of disease with a primary diagnosis or on clinical suspicion. newlineFocus on melanoma detection, irrespective of its unusual occurrence, is that it newlineis often misdiagnosed for other skin malignancies leading to medical newlinenegligence. Sometimes melanoma is detected only when the metastasis has newlineentered the bloodstream or lymph nodes. So effective computational strategies newlinefor early detection of melanoma are essential. There are four principle types newlineof skin melanoma with two sub types: Superficial spreading, nodular, lentigo, newlinelentigo maligna, Acral lentiginous, and Subungual melanoma. Amelanotic newlinemelanoma, one particular type of melanoma, exists in all kinds of skin tones. newlineClassifications of melanoma with its classes are focused on in this research. newlineThis thesis focuses on utilizing ML/DL learning techniques in newlinemelanoma detection. Improvement in image quality is achieved using newlinedeconvolution techniques. Both blind and non-blind image deconvolution newlineapproaches are investigated here. Optimized blind image deblurring is done newlineusing the probabilistic latent semantic analysis technique. A novel approach newlinenamed the ADGMM model is used where descriptors of the input image are newlineretrieved using GMM and fed to an autoencoder to retrieve the relevant newlineinformation missing in the image leading to a sharp output. These descriptors newlineare used for feature enhancement and dimensionality reduction in convolution newlineneural networks. This method also aids in reducing the reconstruction error newlineand provides a quality image. ADGMM portrays better execution conversely newlinethan the best-in-class strategies of the accessible datasets. It gives promising newlineresults and can be reached out with the adjustment of some other CNN layers. newlinePerformance of the deconvolution approaches using Wiener filtering newline