Skin Lesion Classification Using CNNs With Patch-Based Attention and Diagnosis-Guided Loss Weighting

Area of Science:

• Dermatology
• Computer Vision
• Medical Imaging Analysis

Background:

• Effective skin lesion classification is crucial for early diagnosis.
• High-resolution images and imbalanced datasets pose significant challenges for current classification models.

Purpose of the Study:

• To develop methods for effective skin lesion classification using high-resolution images.
• To address the challenge of class imbalance in multi-class skin lesion datasets.

Main Methods:

• A novel patch-based attention architecture was proposed to integrate high-resolution image patches.
• Performance of modified pretrained architectures with patch-based attention was evaluated.
• Class imbalance was addressed using oversampling, balanced batch sampling, and class-specific loss weighting.
• A novel diagnosis-guided loss weighting method was introduced.

Main Results:

• The patch-based attention mechanism improved mean sensitivity compared to previous methods.
• Class balancing techniques significantly enhanced mean sensitivity.
• The diagnosis-guided loss weighting method outperformed standard loss balancing methods.

Conclusions:

• The proposed patch-based attention mechanism effectively utilizes high-resolution images with pretrained architectures without downsampling.
• The diagnosis-guided loss weighting method provides an effective solution for training models with imbalanced skin lesion data.
• These methods enhance automatic skin lesion classification and are applicable to other clinical imaging tasks.