Multi Class Classification of Melanoma in Dermoscopic Images Using Deep Convolutional Neural Network

Abstract

Melanoma is a type of skin cancer and it is characterized from the experts as the most aggressive. An early diagnosis and a surgery removal can give to the patient almost 99% survival rate. Several Computer-Aided Diagnosis (CAD) systems have been proposed to assist dermatologists in an early diagnosis. This thesis, is dealing with the processing of colour images that depict images of patients with possible melanoma. The main point is to build a system to identify cases that could be potentially dangerous. The proposed system performs (i), Preprocessing using four filters namely, Mean Filter, Median Filter, Wiener Filter and Gaussian Filter (ii), Melanoma Segmentation using Triclass Thresholding Method (TTM) (iii), ABCD rule based feature extraction and Machine Learning (ML) based classification and (iv) Deep Convolutional Neural Network (DCNN) based classification. newlineThis research is estimated using two diand#64256;erent publically free available databases namely, PH2 and ISBI 2016 challenge database. The PH2 database includes dermoscopic images from Pedro Hispano Hospital. The dermoscopic images are extracted using a Turbingen Mole Analyser System [TMAS].The eight-bit RGB colour images are enlarged at a resolution of newline768 × 560 pixels. The database consists of 200 images, which includes 160 non-melanoma images and 40 melanoma images. According to the proficient dermatologist, the medical descriptive notes and ground truth for every image are included. The ISBI 2016 database consists of complex image that includes several artifacts such as air bubbles, hair, and several colours. These are challenge database consists of eight-bit RGB colour images at a resolution of 540 × 722 pixels. The database consists of 1279 images, that includes 248 melanoma images and 1031 non-melanoma images (Benign). The pairing of the original images with the ground truth is noted by the professional dermatologists. The 200 images from PH2 database is used in preprocessing, segmentation, feature extraction and ML based classification. newline