Self-supervised multi-modality learning for multi-label skin lesion classification.

Area of Science:

• Artificial Intelligence in Dermatology
• Medical Image Analysis
• Computational Pathology

Background:

• Clinical diagnosis of skin lesions relies on dermoscopic and clinical images.
• Supervised deep learning models require large labeled datasets, which are costly and time-consuming to acquire.
• Existing methods struggle with multi-attribute annotation for skin lesion classification.

Purpose of the Study:

• To develop a self-supervised learning (SSL) algorithm for multi-modality, multi-label skin lesion classification.
• To reduce the dependency on large labeled datasets in skin lesion diagnosis.
• To improve the accuracy of melanoma diagnosis and lesion attribute identification.

Main Methods:

• Proposed a multi-modality SSL algorithm that maximizes similarities between paired dermoscopic and clinical images.
• Introduced a novel multi-modal, multi-label SSL strategy generating pseudo-labels via clustering.
• Developed a label-relation-aware module to refine pseudo-label embeddings and capture attribute interrelationships.

Main Results:

• The algorithm was validated on the seven-point skin lesion dataset, outperforming state-of-the-art SSL methods.
• Demonstrated significant improvements in area under the ROC curve, precision, sensitivity, and specificity.
• Observed enhanced performance across various lesion attributes and melanoma diagnoses.

Conclusions:

• The developed SSL algorithm provides an efficient solution for multi-modality, multi-label skin lesion classification.
• Effectively leverages complementary information from dermoscopic and clinical images and attribute interrelationships.
• Holds potential for improving the accuracy of clinical diagnosis in dermatology.