Deep Learning System in Curvelet Domain for Skin Cancer Diagnosis

Early identification and diagnosis are essential in order to provide appropriate treatment and reduce the mortality rate due to skin cancer. However, Skin Cancer Diagnosis (SCD) may be difficult, even for experienced doctors, since it needs visual inspection of skin lesions and a high degree of knowledge to differentiate between anomalies. This makes it difficult to detect skin cancer. Computer Aided Diagnosis (CAD) systems have the potential to improve the accuracy of SCD. This study proposes a Curvelet based Deep Learning Approach (C-DLA) for SCD. At first, the dermoscopic images are decomposed using Curvelet transform. Then the low frequency components are fed to the Convolutional Neural Network (CNN) for the classification. Results of the proposed C-DLA approach on PH2 database images shows promising results of 99% accuracy for the classification. The proposed CDLA can potentially provide more accurate diagnoses of skin lesions, as it can analyze deep features from the convolutional layers and detect patterns by dense layer that may be difficult for human observers to discern.