Automatic glaucoma detection based on transfer induced attention network.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible vision loss.
• Automatic glaucoma detection from fundus images is crucial but limited by data requirements.
• Existing methods struggle with insufficient labeled data for robust model training.

Purpose of the Study:

• To develop an effective glaucoma detection model using transfer learning.
• To address the challenge of limited labeled data in glaucoma diagnosis.
• To leverage features from similar ophthalmic datasets for improved diagnostic performance.

Main Methods:

• Proposed a Transfer Induced Attention Network (TIA-Net) for glaucoma detection.
• Employed transfer learning from similar ophthalmic datasets.
• Integrated channel-wise attention and maximum mean discrepancy for enhanced feature transferability.

Main Results:

• TIA-Net achieved high accuracy (85.7%/76.6%), sensitivity (84.9%/75.3%), specificity (86.9%/77.2%), and AUC (0.929/0.835) on two clinical datasets.
• Demonstrated superior performance compared to state-of-the-art methods.
• Provided interpretable visualizations of key features.

Conclusions:

• TIA-Net, an attention-based deep transfer learning model, outperforms existing methods for glaucoma diagnosis.
• Leveraging ophthalmic datasets for feature transfer is effective for limited-supervision learning.
• The model shows potential for early diagnosis in other medical applications.