GraVIS

GraVIS enhances self-supervised representation learning for dermatology image analysis by grouping augmented views and using a triplet-loss-inspired objective with a hardness-aware attention mechanism to learn robust, invariant features for lesion segmentation and disease classification.

Key Features:

• Triplet Loss Optimization: Applies a triplet-loss-inspired objective to group homogeneous dermatology image views and separate heterogeneous ones to learn robust self-supervised features.
• Hardness-Aware Attention Mechanism: Integrates a hardness-aware attention that prioritizes homogeneous augmented views with similar appearance to improve representation quality.
• Mitigation of NCE Limitations: Addresses limitations of noise-contrastive estimation (NCE) that rely on single homogeneous image pairs by grouping multiple augmented views.
• Empirical Performance Gains: Demonstrates outperforming transfer learning and other self-supervised methods in lesion segmentation and disease classification, with improvements up to 5% under extremely limited supervision.
• Competitive Pretrained Representations: Pre-trained GraVIS weights enable a single model to surpass ensemble strategies, as demonstrated on the ISIC 2017 challenge.

Scientific Applications:

• Unannotated dermatology image representation learning: Learns transferable representations from unannotated dermatology image datasets for downstream tasks.
• Disease classification: Improves feature representations for supervised and limited-supervision disease classification tasks.
• Lesion segmentation: Provides learned features that improve lesion segmentation performance under limited annotation regimes.

Methodology:

Grouping augmented views of dermatological images from independent sources, optimizing a triplet-loss-inspired objective, and applying a hardness-aware attention mechanism to mitigate NCE single-pair limitations.