MAG-SD

MAG-SD classifies pneumonia, including COVID-19, from chest X-ray (CXR) images using a multiscale attention-guided deep network with soft distance regularization.

Key Features:

• Multiscale Attention Mechanism (MA-Net): MA-Net generates prediction vectors and attention maps from multiscale feature representations to capture spatial information across scales.
• Attention-Guided Augmentations: Attention-guided augmentations produce meaningful variations in CXR training data to enhance model robustness and mitigate limited-data issues.
• Soft Distance Regularization: Soft distance regularization reduces noise and enhances feature discrimination by balancing feature similarity and diversity during learning.

Scientific Applications:

• Clinical diagnosis and triage: Automated classification of CXR images to support rapid and accurate identification of COVID-19 and other pneumonia cases in clinical settings.
• Pandemic response with limited data: Robust screening and case identification using CXR datasets when labeled data are scarce during emerging outbreaks.

Methodology:

MAG-SD uses a Multiscale Attention Network (MA-Net) to produce prediction vectors and attention maps from multiscale feature representations, applies attention-guided augmentations to training data, and employs soft distance regularization to refine feature discrimination.