SkinLiTE: Lightweight Supervised Contrastive Learning Model for Enhanced Skin Lesion Detection and Disease...

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Medical Imaging
• Dermatological Diagnostics

Background:

• Skin lesion detection and classification are crucial for early diagnosis and treatment of skin cancer.
• Existing AI models often struggle with the complexity and imbalance of skin lesion datasets.
• Need for efficient and accurate AI tools in dermatological image analysis.

Purpose of the Study:

• Introduce SkinLiTE, a lightweight supervised contrastive learning model for enhanced skin lesion detection and typification.
• Leverage labeled data to learn generalizable representations for improved performance.
• Address the challenges of complexity and imbalance in skin lesion datasets.

Main Methods:

• Employs a two-phase learning process: contrastive learning for feature representation and supervised classification.
• Utilizes an encoder network and projection head with contrastive loss to minimize intra-class variations and maximize inter-class differences.
• Evaluated on three datasets from the Skin Cancer ISIC 2019-2020 challenge.

Main Results:

• SkinLiTE achieved superior performance in accuracy, AUC, and F1 scores compared to traditional supervised models.
• Demonstrated high accuracy (0.9087) for binary skin lesion classification with AugMix augmentation.
• Showcased comparable results to state-of-the-art approaches without external data, highlighting efficiency.

Conclusions:

• SkinLiTE represents a significant advancement in dermatological AI, offering a robust and accurate tool for skin lesion analysis.
• Its lightweight architecture and ability to handle imbalanced data make it suitable for Internet of Medical Things (IoMT) applications.
• Contributes to AI in healthcare by showcasing innovative learning methodologies for medical image analysis.