DACB-Net: Dual Attention Guided Compact Bilinear Convolution Neural Network for Skin Disease Classification

TL;DR

The paper tackles skin disease classification under data-scarce, imbalanced conditions by introducing DACB-Net, a three-branch architecture that combines Dual Attention Mechanisms with compact bilinear pooling. It features two embedding branches supervised by image-level labels and a global/local attention design guided by Complement Cross Entropy to improve both accuracy and interpretability. Key contributions include the novel Dual Attention Mechanism (DAM) that fuses Channel and Spatial attention, the use of non-homologous BCNN with compact pooling to manage high-dimensional bilinear features, and a loss that leverages incorrect class information to mitigate imbalance. Empirical results on HAM10000 and ISIC2019 show state-of-the-art performance with strong robustness and interpretable Attention Heat Maps, suggesting practical utility in clinical skin disease diagnosis while reducing computational demands.

Abstract

This paper introduces the three-branch Dual Attention-Guided Compact Bilinear CNN (DACB-Net) by focusing on learning from disease-specific regions to enhance accuracy and alignment. A global branch compensates for lost discriminative features, generating Attention Heat Maps (AHM) for relevant cropped regions. Finally, the last pooling layers of global and local branches are concatenated for fine-tuning, which offers a comprehensive solution to the challenges posed by skin disease diagnosis. Although current CNNs employ Stochastic Gradient Descent (SGD) for discriminative feature learning, using distinct pairs of local image patches to compute gradients and incorporating a modulation factor in the loss for focusing on complex data during training. However, this approach can lead to dataset imbalance, weight adjustments, and vulnerability to overfitting. The proposed solution combines two supervision branches and a novel loss function to address these issues, enhancing performance and interpretability. The framework integrates data augmentation, transfer learning, and fine-tuning to tackle data imbalance to improve classification performance, and reduce computational costs. Simulations on the HAM10000 and ISIC2019 datasets demonstrate the effectiveness of this approach, showcasing a 2.59% increase in accuracy compared to the state-of-the-art.

Figures (18)

• Figure 1: Improved CAM mechanism.
• Figure 2: Improved SAM mechanism.
• Figure 3: Structure of dual attention mechanism.
• Figure 4: Bilinear attention mechanism model.
• Figure 5: Proposed DACB-Net architecture.