부산대학교 도서관

Skin cancer has a high fatality rate, especially in Western countries. Early detection of skin cancer prolongs human life and is helpful to cure disease. Dermoscopy inspection is a frequently utilized noninvasive method to diagnose skin cancer. Visual inspection of dermoscopy images takes more inspection time, and the decision is based on the individual perception of dermatologists. The existing methods for skin cancer classification utilize only spatial information. However, the spectral domains of information are missing to classify skin lesions. Therefore, the performance of the existing models is moderate. To improve the performance of skin cancer classification, this work proposed novel hand-crafted features formulated using image-, spectrogram-, and cepstrum-domain features. The developed hand-crafted features use spatial as well as spectral information. Furthermore, the developed hand-crafted features are given as input to a newly developed 1-D multiheaded convolutional neural network (CNN) for the classification of skin lesions, using the challenging HAM10000 and Dermnet datasets. The performance of the proposed network is compared with the other existing state-of-the-art methods on the same dataset. From the experimental analysis, the proposed network achieved an accuracy of 89.71% on the HAM10000 dataset and an accuracy of 88.57% on the Dermnet dataset. The proposed method may be used to enhance the performance of clinical diagnosis measurement.