Abstract:
In recent years, the occurrence and mortality rate due to skin cancer has increased to a
higher extent worldwide. It is crucial to identify such cancers early and accurately to
provide proper treatment, and research has shown that deep intelligent learning-based ways
to address this issue have been proved successful. The main motivation of this study is to
classify skin cancer using deep learning techniques on dermoscopy dataset with optimal
performance while training time taken into account. The aim is to develop such an
automated framework which can perform optimally across three different dermoscopy
datasets having diverse characteristics. We have proposed a model SkinNet-14 by altering
compact convolutional transformer (CCT) using 32 × 32 sized input image which results
in minimizing time complexity to classify skin cancer into different classes. The SkinNet-
14 architecture is developed through ablation study conducted on CCT model using HAM
dataset. Prior to that, several data augmentation techniques and preprocessing methods are
applied to enhance the image quality and quantity of all the datasets. Afterwards, the
proposed model is evaluated with the rest two datasets. Results show that, the model which
was proposed, achieved an accuracy of 97.85% on the HAM dataset, 96.0% on the ISIC
dataset, and 98.14% on the PAD dataset. Moreover, the proposed model yields better
performance in terms of number of parameters, accuracy and training time than six transfer
learning model while training with 32 × 32 sized images.
