TailBoost: Tail-Synthetic Learning for Boosting Long-Tailed Skin Cancer Image Classification

Area of Science:

• Dermatology
• Medical Imaging
• Computer Science

Background:

• Skin cancer image datasets often have imbalanced class distributions, with few dominant classes and many underrepresented 'tail' classes.
• This imbalance negatively impacts deep learning model performance in classifying skin lesions.
• Existing methods like traditional mixup can distort important diagnostic features by not focusing on regions of interest.

Purpose of the Study:

• To develop a novel framework, TailBoost, to improve long-tailed skin cancer image classification.
• To address the limitations of traditional mixup techniques in preserving diagnostic features.
• To enhance the performance of deep learning models on imbalanced skin cancer datasets.

Main Methods:

• Introduced the TailBoost framework utilizing a novel strategy called SPMix.
• SPMix generates synthetic tail-class images by combining tail-class and head-class images guided by saliency maps.
• Incorporated supervised contrastive learning with class-center rebalance to refine learned representations.

Main Results:

• TailBoost demonstrated superior performance compared to existing state-of-the-art long-tailed learning methods.
• Experiments were conducted on ISIC2018, ISIC2019, and PAD-UFES-20 datasets.
• The SPMix strategy effectively preserved and enhanced discriminative features of tail-class images.

Conclusions:

• TailBoost offers an effective solution for long-tailed skin cancer image classification.
• The proposed SPMix method successfully mitigates feature distortion in synthetic data.
• This framework significantly improves the accuracy of deep learning models in identifying rare skin cancer types.