DeepGrading: Deep Learning Grading of Corneal Nerve Tortuosity.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Corneal nerve fiber tortuosity assessment is crucial for diagnosing eye diseases.
• Current manual grading methods lack standardization and accuracy.
• Automated quantification is needed for reliable clinical decision-making.

Purpose of the Study:

• To develop a fully automated deep learning method for grading corneal nerve fiber tortuosity.
• To improve the accuracy and interpretability of tortuosity grading using corneal confocal microscopy (CCM) images.
• To establish a robust tool for disease understanding and clinical application.

Main Methods:

• A two-stage deep learning approach utilizing a fine-tuned feature extraction backbone with a novel bilinear attention (BA) module.
• The BA module captures long-range dependencies and global contexts of nerve fibers.
• An auxiliary tortuosity grading network (AuxNet) refines grading, with results fused for final output.

Main Results:

• The proposed method achieved an overall accuracy of 85.64% in a four-level classification task.
• Demonstrated significant differences in tortuosity levels between healthy controls and diabetic patients.
• Outperformed existing methods in tortuosity grading accuracy.

Conclusions:

• The automated deep learning method provides accurate and interpretable grading of corneal nerve tortuosity.
• This technique has potential for early disease detection and monitoring, especially in diabetic patients.
• A publicly available dataset and code are provided to facilitate further research.