Automatic Segmentation and Classification of Skin Lesions in Dermoscopic Images

Adil Humayun, Khan (2024) Automatic Segmentation and Classification of Skin Lesions in Dermoscopic Images. PhD thesis, UNIMAS.

	PDF Thesis PhD_ Adil Humayun Khan - 24 pages.pdf Download (4MB)
	PDF (Please get the password by email to repository@unimas.my , or call ext: 3942 / 3973 / 3933) Thesis PhD_ Adil Humayun Khan.ftext.pdf Restricted to Registered users only Download (4MB) | Request a copy
	PDF Thesis PhD_ Adil Humayun Khan.dsva.pdf Restricted to Repository staff only Download (270kB) | Request a copy

Abstract

Skin cancer has a malignant type that is Melanoma and two benign types Nevus and Seborrheic Keratosis (SK). Diagnosis of actual skin cancer type is very challenging for physicians because there is no fixed colour, area, size, or boundary associated with a specific type. There is a remarkable resemblance between different types of skin cancer. Therefore, a computer-based process is required to help physicians in diagnosing actual cancer types timely. This process consists of four steps: pre-processing, segmentation, feature extraction, and classification. In this research, we evaluate proposed algorithms on two datasets, International Skin Imaging Collaboration (ISIC) and PH2 (Dermatology Service of Hospital Pedro Hispano, Matosinhos, Portugal). Statistical techniques such as Anisotropic Diffusion Filter (ADF), local contrast enhancement, and haze reduction in the CIELAB colour space, are incorporated in the proposed algorithms for the pre-processing step. Hair removal is implemented through black-hat morphological processing and total variation based inpainting. For segmentation, the first proposed algorithm is based on the boundary condition model, which is tested over the ISIC dataset and achieved 96% of accuracy. The second segmentation algorithm combines Delaunay triangulation clustering in the spatial domain and Particle Swarm Optimization (PSO). This proposed algorithm achieved segmentation accuracy of 96.8% and 92.1% for ISIC and PH2 datasets respectively. The third proposed segmentation algorithm involves two pipelines for feature extraction: split & merge methods and Contextual Encoding Network (EncNet) with an attention mechanism. This proposed algorithm achieved segmentation accuracy of 97.8% and 96.7% for ISIC and PH2 datasets respectively. The first proposed classification algorithm utilizes a Convolution Neural Network (CNN), in which the number of parameters and layers are reduced significantly, and 96% of classification accuracy is achieved on the ISIC dataset. The second proposed iv classifier extracts statistical features, and a boost ensemble learning technique is implemented using Support Vector Machines (SVM) as initial classifiers and Artificial Neural Networks (ANN) as the final classifier. This proposed classifier achieved 97.9% classification accuracy on the ISIC dataset. In the third classification algorithm, hybrid features are extracted using AlexNet and VGG-16 through a transfer learning approach where parameter manipulation is implemented to simplify the network. This proposed classifier achieved 98.2% classification accuracy on the ISIC dataset. These algorithms are proposed while implying modifications to existing statistical, machine, and deep learning methods.

Actions (For repository members only: login required)

View Item