GAN

GAN implements fixed-point image-to-image translation to remove artifacts and enable weakly-supervised disease detection and localization in medical images.

Key Features:

• Fixed-Point Translation: Identifies and manipulates the minimal subset of target pixels required for domain translation, enabling "virtual healing" by transforming images to a healthy state and revealing diseased regions via image subtraction.
• Training Methodology: Supervises same-domain translation with a conditional identity loss and regularizes cross-domain translation with revised adversarial, domain classification, and cycle consistency losses.
• Disease Detection Framework: Derives a weakly-supervised framework for disease detection and localization that uses only image-level annotations to localize pathology without pixel-level labels.
• Multi-domain Object Removal and Integrity Preservation: Performs multi-domain image-to-image translation and object removal while preserving image integrity and reducing artifacts.

Scientific Applications:

• Medical Imaging : Artifact removal and highlighting of diseased regions through virtual healing and image subtraction to support diagnostic interpretation.
• Weakly-Supervised Localization: Disease localization using only image-level annotations, demonstrating improved accuracy and efficiency compared to existing methods.
• Multi-Domain Translation Research: Development and evaluation of multi-domain image-to-image translation approaches for tasks requiring minimal pixel interventions.

Methodology:

Implements fixed-point translation to minimize the subset of target pixels for transformation; supervises same-domain translation with a conditional identity loss; regularizes cross-domain translation with revised adversarial, domain classification, and cycle consistency losses; and uses virtual healing plus image subtraction to derive weakly-supervised disease localization from image-level annotations.