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

Although melanoma is one of the deadliest types of cancer, its early detection can significantly improve the patient's chance of survival. Biopsy is the most accurate diagnostic method, but due to its invasiveness, non-invasive diagnostic techniques such as dermoscopy have become highly attractive. Dermoscopy enables the identification of melanoma by analyzing images captured from the skin. However, many lesions may appear suspicious despite being benign, which makes accurate classification challenging. In this study, the efficiency of deep learning methods in multiclass classification of dermoscopic images is evaluated. Two well-known datasets, ISIC۲۰۱۹ and HAM۱۰۰۰۰, containing ۸ and ۷ lesion types respectively, were used to assess the detection performance of convolutional neural networks (CNNs). The results show that deep learning achieves significantly higher accuracy than traditional approaches. Furthermore, a new data augmentation technique based on merging multiclass datasets is proposed, which enhances CNN performance and improves lesion recognition accuracy in dermoscopy images.

کلیدواژه ها:

Skin cancer ، Melanoma ، Deep learning ، Dataset fusion ، Data augmentation ، Multiclass classification