Dynamic feature splicing for few-shot rare disease diagnosis

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Acquiring annotated medical images for rare disease diagnosis is difficult.
• Few-shot learning (FSL) aims to identify rare diseases with limited data.
• Existing FSL methods struggle with complex medical image characteristics and high intra-class variance.

Purpose of the Study:

• To propose a novel dynamic feature splicing (DNFS) framework for few-shot rare disease diagnosis.
• To enhance both low-level and high-level features of rare disease classes using knowledge from abundant base classes.
• To improve the accuracy and performance of rare disease diagnosis in a few-shot setting.

Main Methods:

• Developed a dynamic feature splicing (DNFS) framework incorporating position coherent (P-DNFS) and semantic coherent (S-DNFS) modules.
• P-DNFS utilizes a lesion-oriented Transformer to detect lesion regions for low-level feature splicing.
• S-DNFS explores cross-image channel relations for high-level feature splicing based on semantic consistency.

Main Results:

• The DNFS framework dynamically and iteratively enriches both low-level and high-level features.
• Generated abundant spliced features, leading to a more accurate decision boundary.
• Demonstrated superior performance compared to state-of-the-art approaches on three medical image classification datasets.

Conclusions:

• The proposed DNFS framework effectively addresses the challenges of few-shot rare disease diagnosis.
• Dynamic feature splicing significantly improves diagnostic accuracy by leveraging knowledge from base classes.
• DNFS offers a promising approach for rare disease identification in resource-limited scenarios.