Multi-level contexts aggregation for melanoma recognition under feature confusion regularization

Due to variation in the appearance of the skin lesions, the noisy background, the small training dataset, and label imbalance, it is still a challenging task for melanoma recognition in dermoscopy images. Existing deep convolution network-based approaches usually adopted the global context captured in the fully connection layers (FC) to recognize melanoma, ignoring the local contexts which are beneficial for melanoma recognition. We adopted the DenseNet without FC to keep the maximum multi-scale skin lesion features flowing between layers. An attention module is designed to focus on the local contexts of the skin lesion captured in the intermediate Dense Blocks (InterDBs). After aggregating the local contexts and the global context captured in the last Dense Block (LastDB), the performance of melanoma recognition is improved. We proposed a novel feature confusion regularization, which reduces the Euclidean distance (L₂) between features of the pair inputs with different labels to prevent the proposed network from extracting excessive discriminative features, specific for the samples in the small training set, and avoid overfitting. We trained the proposed network on the ISIC 2017 without using extra data and tested the robustness of it on the ISIC 2016, ISIC 2019. The experimental results show that our network outperforms the state-of-the-art studies, especially for average precision (0.694) and sensitivity (0.693), which are improved by 2.2 and 3.5%, respectively. Evaluation on the ISIC 2019 indicates our network improves by 8%, 3% for average precision and average sensitivity, respectively, which shows it is robust and suitable for clinical application.