AUTOMATIC SKIN LESION CLASSIFICATION BASED ON CXception.

Automatic classification of skin lesions remains a challenging task due to the insufficient training data, the morphological diversity of skin lesions, and the existence of artefacts and intrinsic cutaneous features in dermoscopy images. We propose a novel convolutional neural network termed CXception to tackle these challenges. CXception is constructed by plugging a coordinate attention (CA) block into the basic Xception architecture. CA block is a simple yet efficient attention mechanism, which can encode both inter-channel relationships and long-range dependencies that preserve precise positional information. Integrating CA into Xception enables the model to learn more expressive representations, hence improving the diagnostic performance of skin lesions effectively and significantly. The proposed method is supposed to handle the outliers that are not in the training set. We tackle this problem by using an efficient data-driven approach. Besides, this multiclassification task comes with the problem of heavy class imbalance. We deal with this issue by adopting an optimized loss function called the class-weighted cross-entropy loss. The experimental results on the public benchmark dataset (ISIC 2019 dataset) demonstrate the superior performance of the proposed method relative to that of the baselines (backbone network and classical classification models) and state-of-the-art approaches. Code of the proposed method is available at https://github.com/shanpufang/skin-lesion-nine. [ABSTRACT FROM AUTHOR]