Light Weight CNN Model for Accurate and Efficient Skin Cancer Classification

This paper, proposes a lightweight model: a modified Convolutional Neural Network (CNN) to perform the multi-class skin cancer classification with the HAM10000 dermoscopic image dataset. The architecture of the proposed model is capable of detecting seven skin lesion categories through computationally efficient operations including: separable convolutions, Swish activation, spatial dropout, and global average pooling. These architectural improvements significantly speed up the model convergence and legitimate the number of parameters but it still performs on high scale. The data is pre-processed using median filter,
standardization, and augmentation to handle the class
imbalances. Experiments show that by training in 220
seconds, DCN can achieve 95% accuracy, surpassing several
popular benchmarks of deep learning such as InceptionV4
and ResNet50 in terms of both time and space efficiency.
This renders the model to be very applicable for real-time
diagnosis at healthcare facilities with mobile or limited
resources.