Deep attention branch networks for skin lesion classification

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin lesion classification is challenging due to subtle differences and small lesion sizes in dermoscopy images.
• Existing methods struggle to focus on discriminative lesion regions, limiting classification accuracy.
• Class Activation Mapping (CAM) highlights important areas but isn't available during standard network training.

Purpose of the Study:

• To develop an accurate skin lesion classification model that can focus on semantically meaningful parts of lesions.
• To introduce a novel network architecture capable of generating attention maps during training.
• To address the challenge of class imbalance in skin lesion datasets.

Main Methods:

• Proposed a Deep Attention Branch Network (DABN) integrating attention branches into Deep Convolutional Neural Networks (DCNNs).
• DABN generates Class Activation Maps (CAMs) during training, used as attention maps to guide the network.
• Introduced an Entropy-guided Loss Weighting (ELW) strategy to mitigate the impact of class imbalance.

Main Results:

• Achieved an Average Precision (AP) of 0.719 on the ISIC-2016 dataset and an Area Under the ROC Curve (AUC) of 0.922 on the ISIC-2017 dataset.
• Outperformed state-of-the-art methods without external data or ensemble learning.
• Demonstrated improved mean sensitivity by over 2.6% across different DCNN structures and applicability to multi-class classification.

Conclusions:

• The DABN model adaptively focuses on discriminative regions in dermoscopy images for improved classification.
• The ELW strategy enables effective training despite class imbalance, enhancing overall performance.
• The proposed approach shows potential for broader clinical applications beyond skin lesion classification.