Efficiency Analysis of Convolutional Neural Netwoks in Multiclass Recognition of Skin Lesions via Dermoscopy Images

Although melanoma is considered as one of the deadliest types of cancer, however, detecting this disease in its early stages may lead to saving the patient's life. Biopsy is the most accurate method to diagnose this type of cancer, but due to its invasiveness trying to use non-invasive diagnostic modalities has always been of interest to researchers in this field. Dermoscopy is one of the common non-invasive modalities, in which it is determined whether the patient is healthy or cancerous by analyzing the images captured from the patient's skin. However, there may be several lesions on the skin that, despite their appearance, most of them are not cancerous and even harmless. In this way, recognizing melanoma lesions is the most important challenge in the field of analyzing dermoscopic images especially when it is desired to determine the type of other lesions in this classification process. In this article, the performance of the deep learning paradigm in the classification of multi-type lesions in dermoscopic images is evaluated and the challenges facing deep neural networks in this domain are analyzed. For this purpose, two well-known datasets of dermoscopy images known as ISIC2019 and HAM10000, which have 8 and 7 types of labeled lesions, respectively, were selected and the detection rate of lesions in them was evaluated by applying deep neural networks. The obtained results indicate that although the results obtained from the application of deep learning in the separation of lesions in both of examined datasets are significantly better than the conventional methods, but the resultant accuracy is highly dependent on the volume of data and balance between several classes of lesions. In this way, the use of data augmentation and data balancing methods should be considered as a priority for conducting any research in order to improve the analysis of dermoscopic images by using deep learning.