DH-OOD: A decoupled hybrid framework for robust skin lesion classification via semantic-structural fusion

Area of Science:

• Dermatology
• Computer Vision
• Machine Learning

Background:

• Real-world skin lesion classification faces challenges like imbalanced data, high intra-class variability, and the need to identify out-of-distribution (OOD) samples.
• Traditional monolithic models struggle to address these issues concurrently, limiting their effectiveness in clinical applications.

Purpose of the Study:

• To develop a robust framework for skin lesion classification that addresses class imbalance and OOD sample detection.
• To improve the accuracy and reliability of automated diagnostic tools for dermatological conditions.

Main Methods:

• A multi-stage decoupled hybrid framework integrating Supervised Contrastive Learning (SupCon) for representation learning and a Convolutional Autoencoder (CAE) for structural reconstruction.
• SupCon was employed to create a more balanced feature space, mitigating feature degradation in long-tailed distributions.
• Contrastive semantic features were fused with structural anomaly scores from the CAE for open-set recognition, enabling both classification and rejection of unknown samples.

Main Results:

• The proposed framework achieved a Balanced Accuracy of 78.5% on known skin lesion classes within the ISIC 2019 dataset.
• An improved unknown-class F1-score of 51.3% was recorded, demonstrating effective rejection of OOD samples.
• Semantic-structural fusion enhanced model robustness under long-tailed and open-set conditions.

Conclusions:

• The decoupled hybrid framework effectively addresses key challenges in real-world skin lesion classification, including data imbalance and OOD detection.
• Integrating supervised contrastive learning with structural reconstruction provides complementary signals for improved classification and sample rejection.
• This approach offers a more robust and reliable solution for automated dermatological analysis.