From single to multiple: Generalized detection of Covid-19 under limited classes samples

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning for medical diagnosis requires extensive labeled data and medical expertise, posing challenges in clinical practice.
• Current methods struggle with data scarcity and the complexity of disease representation, limiting their real-world applicability.
• The COVID-19 pandemic highlighted the need for rapid, accurate, and adaptable diagnostic tools.

Purpose of the Study:

• To address the limitations of existing deep learning models in medical image diagnosis.
• To propose a novel framework, Open-Set Single-Domain Generalization for Medical Image Diagnosis (OSSDG-MID), for training models on a single source domain.
• To enable accurate classification of target domain samples and identification of 'unknown' categories outside the source domain's scope.

Main Methods:

• Introduced the Open-Set Single-Domain Generalization for Medical Image Diagnosis (OSSDG-MID) problem setting.
• Developed the Multiple Cross-Matching (MCM) method to generate auxiliary samples outside the source domain's category space.
• Evaluated the approach on diverse cross-domain image classification tasks.

Main Results:

• The proposed MCM method significantly improved the identification of 'unknown' categories.
• OSSDG-MID demonstrated superior performance in single-domain generalization compared to existing methods.
• The approach achieved high accuracy in open-set image classification tasks.

Conclusions:

• The MCM method effectively enhances the ability of deep learning models to generalize to new domains while identifying out-of-distribution samples.
• OSSDG-MID offers a promising solution for developing robust medical image diagnostic tools with reduced reliance on extensive labeled data.
• This research contributes to advancing AI in healthcare by enabling more adaptable and accurate diagnostic systems.