CAA-Net

CAA-Net classifies corneal photographs to perform automated multi-class diagnosis of infectious keratitis, distinguishing normal, viral, fungal, and bacterial presentations.

Key Features:

• End-to-end fully convolutional architecture: Uses a fully convolutional network trained end-to-end for image-based classification of corneal photographs.
• Class-aware classification module: Incorporates a class-aware module trained to extract discriminative features specific to each keratitis type.
• Separate class branches: Employs separate branches for each class (normal, viral, fungal, bacterial) to learn distinct class-specific characteristics.
• Feature integration into main branch: Integrates class-specific discriminative features into a main branch for final decision-making.
• Two attention strategies: Fuses class-specific features with other feature maps using two attention strategies to emphasize relevant features and suppress irrelevant ones.
• Dataset validation: Validated on a dataset of 1886 corneal photographs from 519 patients to assess diagnostic performance.

Scientific Applications:

• Automated diagnosis of infectious keratitis: Supports automated identification of infectious keratitis from corneal photographs for clinical diagnostic workflows.
• Type differentiation: Enables differentiation among normal, viral, fungal, and bacterial keratitis presentations in image-based studies.
• Image-based ophthalmic research: Serves as a model for developing and benchmarking deep-learning approaches in ophthalmic image analysis.

Methodology:

End-to-end fully convolutional network with a class-aware classification module that uses separate class branches whose discriminative features are integrated into a main branch and fused with other feature maps via two attention strategies; trained and evaluated on 1886 corneal photographs from 519 patients.