Cascading size-dependent deep propagation (CADP): Addressing over-smoothing in graph few-shot dermatology...

Area of Science:

• Machine Learning
• Computer Vision
• Medical Image Analysis

Background:

• Graph Neural Networks (GNNs) excel at complex data relationships but suffer from over-smoothing, degrading node representations.
• Over-smoothing homogenizes node features by excessive neighborhood aggregation, limiting GNNs' discriminative power.

Purpose of the Study:

• To introduce Cascading Size-Dependent Deep Propagation (CADP), a novel method to mitigate over-smoothing in graph-based few-shot learning.
• To enhance skin disease classification accuracy using CADP by improving feature representation and label propagation.

Main Methods:

• Constructing graphs using Convolutional Neural Networks (CNNs) for feature extraction from images, with nodes as features and edges as similarity.
• Decoupling feature propagation from neural network transformations to enable deeper information flow and prevent over-smoothing.
• Integrating support labels with query image predictions via Multi-Layer Perceptron (MLP) and optimizing through deep label propagation controlled by size-dependent hyperparameters (K1, K2).

Main Results:

• CADP achieved high accuracy in the 2-way 5-shot setting on dermatology datasets: 78.3% (ISIC 2018), 79.29% (Derm7pt), and 91.92% (SD-198).
• The proposed method demonstrated superior performance compared to existing approaches across all evaluated datasets.

Conclusions:

• CADP effectively mitigates over-smoothing in graph-based few-shot learning for medical image classification.
• The approach shows significant promise for improving diagnostic accuracy in dermatology through enhanced graph-based learning.