Interpreting Deep Learning Studies in Glaucoma: Unresolved Challenges.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning (DL) is rapidly advancing automated image classification in medical specialties like ophthalmology.
• Few DL algorithms have achieved regulatory approval for autonomous diagnosis in specific health conditions.
• International efforts are underway to establish performance benchmarking thresholds for AI in healthcare.

Purpose of the Study:

• To review major deep learning studies in glaucoma diagnosis.
• To identify recurrent challenges and limitations hindering clinical deployment of these algorithms.
• To analyze common issues across studies using fundus photographs, OCT scans, and visual field data.

Main Methods:

• Focused on the three most common input modalities for glaucoma diagnosis: fundus photography, spectral domain optical coherence tomography (SD-OCT), and standard automated perimetry (SAP).
• Analyzed challenges related to algorithm output definition, ground truth dataset reliability, and training dataset representativeness.
• Examined the largest deep learning studies in glaucoma.

Main Results:

• Identified significant challenges in defining glaucoma as an algorithm output.
• Highlighted difficulties in establishing reliable ground truth datasets for training and validation.
• Noted the complexity of compiling representative training datasets that reflect diverse patient populations and disease variations.

Conclusions:

• Widespread clinical deployment of deep learning for glaucoma diagnosis is currently limited by several key challenges.
• Standardization in defining glaucoma, ground truth determination, and dataset representativeness are crucial for advancing AI in ophthalmology.
• Further research and collaborative efforts are needed to overcome these hurdles and enable reliable autonomous diagnosis.