Skin Cancer Diagnosis Based on the Convolutional Neural ?Network: a Comparative Study


  • Jane J. Stephan University of Information Technology and Communications
  • Muhammed Kadhim Hussein Iraqi Commission for Computers and Informatics Informatics Institute for Postgraduate Studies


‎ Skin diseases, technology, classification, features, deep learning, CNN.‎


Skin cancer is one of the leading causes of death in humans, however, it is treatable if caught early. Therefore, early detection of skin cancer contributes to saving many patients. Skin cancer is divided into two types, benign tumor, and malignant tumor that leads to the death of a person if not treated early, and both are similar in appearance only a dermatologist can classify cancer as malignant or benign.

The proposed system consists of several basic stages. The first stage is the creation and provision of a large database, the second stage is the use of data compression techniques (images), and the third stage is the use of artificial intelligence by applying an artificial neural network to image processing technology, specifically in the field of deep learning approach. The database used in the proposed system consists of a set of skin cancer images from the International Skin Imaging Cooperation (ISIC) and a set of images also brought from the Medical City in Iraq (Dermatology Consultation Department). Be clear and free of distortion at a good rate.

Huffman technology is used to compression images while preserving image information from loss, reducing image size, saving storage space, saving time, and thus increasing system speed, as a neural network (deep learning) was used with the SVM classifier for the support machine. Also, a set of deep learning models VGG16, AlexNet, ResNet-50 and Inception v3 were used only without any modifications to the models except the last layer of each model. Finally, a special model that detects skin cancer (SkinNet) was proposed.

The method used in detecting skin cancer is deep learning that works by inserting compressed images and then splitting the images 70% for the training process, 30% for the testing process, and the proposed model (SkinNet) performed better with accuracy, and the performance was with 98.2% accuracy.