Detection of melanoma in dermoscopic images by integrating features extracted using handcrafted and deep learning models

Abstract

Melanoma is amongst the most aggressive form of skin cancer. The manual detection of melanoma using dermoscopic images is a labor-intensive and time-consuming process, which requires a lot of expertise too. This necessitates the development of an intelligent classifier that detects melanoma accurately so that proper and timely treatment can be given to the patient. However, melanoma detection is a challenging task due to the presence of noise like air bubbles, hair, etc. in the dermoscopic images. First, we propose three methods to remove hair in dermoscopic images using various morphological operations. As the quality of features affects the performance of a classifier, we propose to use integrated features that are extracted using handcrafted (HC) feature extraction techniques and deep learning model (DLM) from dermoscopic images, to enhance the performance of the classifier. Two DLMs: ResNet50V2 and EfficientNet-B0 are employed for feature extraction and Artificial Neural Network (ANN) is used for classification. The proposed approach is evaluated using two datasets: the HAM10000 dataset taken from the ISIC 2018 challenge, consisting of 10,015 dermoscopic images belonging to 7 classes, and the PH2 dataset consisting of 200 dermoscopic images with 40 melanoma and 160 non-melanoma images. Experimental results show that the proposed hair removal methods along with the integrated features improve the accuracy of melanoma detection as compared to when (i) no-pre-processing is applied or (ii) HC or deep learning (DL) features are used alone. An accuracy of 94.9% and 98% is achieved on HAM10000 and PH2 datasets respectively when integrated features extracted using HC techniques and EfficientNet-B0 are used for classification along with hair removal techniques. The performance of the proposed approach on both datasets is comparable to the existing state-of-the-art classifiers for melanoma detection.