Benign and Malignant Skin Lesion Classification Comparison for Three Deep-Learning Architectures

Early detection of melanoma, which is a deadly form of skin cancer, is vital for patients. Differential diagnosis of malignant and benign melanoma is a challenging task even for specialist dermatologists. The diagnostic performance of melanoma has significantly improved with the use of images obtained via dermoscopy devices. With the recent advances in medical image processing field, it is possible to improve the dermatological diagnostic performance by using computer-assisted diagnostic systems. For this purpose, various machine learning algorithms are designed and tested to be used in the diagnosis of melanoma. Deep learning models, which have gained popularity in recent years, have been effective in solving image recognition and classification problems. Concurrently with these developments, studies on the classification of dermoscopic images using CNN models are being performed. In this study, the performance of AlexNet, GoogLeNet and Resnet50 CNNs were examined for the classification problem of benign and malignant melanoma cancers on dermoscopic images. Dermoscopic images of 19373 benign and 2197 malignant lesions obtained from ISIC database were used in the experiments. All three CNNs, which were the former winners of the ImageNet competition, have been reconfigured to perform binary classification. In the experiments 80% of the images were used for training and the remaining 20% were used for validation. All experiments were performed with the same parameters for each CNN models. According to the experiments ResNET50 model achieved the best performance with 92.81% classification accuracy and AlexNet was first-ranked in terms of the time complexity measurements. The development of new models based on existing CNN models with a focus on dermoscopic images will be the subject of future studies.