One-shot learning for generalization in medical image classification across modalities.

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Computer Vision

Background:

• Generalizability is a major hurdle for medical sensing technologies, especially with limited or poor-quality imaging data.
• Existing methods struggle to adapt to diverse imaging modalities like X-ray, microscopy, and CT scans.

Purpose of the Study:

• To propose a generalized, robust, and lightweight one-shot learning method for cross-modality medical image classification.
• To address the challenge of limited data by enabling effective training with a single image per class.

Main Methods:

• Introduced a Collaborative One-Shot Training (COST) approach combining meta-learning and metric-learning.
• Employed gradient generalization in dense and fully connected layers for faster convergence.
• Utilized a lightweight Siamese network with triplet loss and shared parameters.

Main Results:

• Achieved an average accuracy of 91.5% and AUC of 0.89 across 12 MedMNIST2D datasets.
• Outperformed state-of-the-art models (ResNet-50, AutoML) by over 10% on specific datasets.
• Demonstrated superior performance on the OCTMNIST dataset (AUC 0.91 vs. ResNet-50's 0.77).

Conclusions:

• The proposed COST method significantly improves convergence speed and accuracy compared to traditional one-shot learning.
• The lightweight architecture (0.15M parameters) is suitable for resource-constrained devices.
• The model offers a generalized and robust solution for medical image classification across various modalities.