A Modified Inception-v4 for Imbalanced Skin Cancer Classification Dataset

Abstract

Deep learning architectures, especially deep convolutional neural networks (CNN) achieve high accuracy on object classification and localization tasks. Achieving such high accuracy requires powerful devices. In this paper, rather than an ensemble of multiple complex models, a single Inception-v4 model is adapted to classify extracted from the HAM10000 dataset. The proposed model is enhanced by employing feature reuse using long residual connection in which the features extracted from earlier layers are concatenated with the high-level layers to increase the model classification performance. The dataset used in this study is imbalanced; therefore, a data sampling approach is used to mitigate the data imbalance effect. The proposed architecture achieves an accuracy of 94.7% using the provided test set at the official benchmark for the International Skin Imaging Collaboration (ISIC) 2018.